Published:
If I reopened the 2026 folder in 2030, I would check one thing first: did the work completed with AI change a real problem and leave evidence that others could verify and build upon?
Published:
假设在 2030 年重新打开 2026 年的文件夹，我会先查一件事：今天借助 AI 完成的工作，是否改变了一个现实问题，并留下别人能够复验和继承的证据。
Published:
| Original source | Tencent Tech; transcript edited | Gu Lingyu; original editors | Xu Qingyang, Su Yang; 2030 annotations and second edit | Koutian Wu |
Published:
原文来源｜腾讯科技；原文整理｜顾翎羽；原文编辑｜徐青阳、苏扬；2030 注释与二次编辑｜Koutian Wu
Published:
My verdict on Air Canada after flying from Singapore to Austin through Vancouver is simple: what a mess.
Published:
这次从新加坡经温哥华飞回奥斯汀，我对 Air Canada 的评价很简单：太拉了。
Published:
My verdict on Air Canada after flying from Singapore to Austin through Vancouver is simple: what a mess.
Published:
这次从新加坡经温哥华飞回奥斯汀，我对 Air Canada 的评价很简单：太拉了。
Published:
Day 110 of Benchmark Radar. The scan fetched 1,548 items, published 676, and marked 352 as recommended.
Published:
Benchmark Radar 第 110 天。你的编程智能体从没见过机器人。本轮抓取 1,548 条，发布 676 条，其中 352 条标为推荐。
Published:
Day 109 of Benchmark Radar. The scan fetched 1,181 items, published 467, and marked 228 as recommended.
Published:
Benchmark Radar 第 109 天。500 项任务里 433 项是占位符。本轮抓取 1,181 条，发布 467 条，其中 228 条标为推荐。
Published:
Day 108 of Benchmark Radar. The scan fetched 1,292 items, published 573, and marked 249 as recommended.
Published:
Benchmark Radar 第 108 天。你的排行榜分不清谁真正赢了。本轮抓取 1,292 条，发布 573 条，其中 249 条标为推荐。
Published:
Day 107 of Benchmark Radar. The scan fetched 1,148 items, published 611, and marked 307 as recommended.
Published:
Benchmark Radar 第 107 天。你的回测早就知道未来。本轮抓取 1,148 条，发布 611 条，其中 307 条标为推荐。
Published:
Day 106 of Benchmark Radar. The scan fetched 1,111 items, published 496, and marked 283 as recommended.
Published:
Benchmark Radar 第 106 天。真正被测的是运行框架。本轮抓取 1,111 条，发布 496 条，其中 283 条标为推荐。
Published:
Day 105 of Benchmark Radar. The scan fetched 1,322 items, published 599, and marked 205 as recommended.
Published:
Benchmark Radar 第 105 天。配置决定了排名。本轮抓取 1,322 条，发布 599 条，其中 205 条标为推荐。
Published:
Your next cloud instance is alive, it needs to be fed every three days, and it will be dead in six months whether you cancel or not.
Published:
你的下一台云服务器是活的，每三天要喂一次饭，而且不管你续不续费，它六个月后都会死。
Published:
Day 104 of Benchmark Radar. The scan fetched 1,462 items, published 744, and marked 307 as recommended.
Published:
Benchmark Radar 第 104 天。评分器会被它评的答案骗过。本轮抓取 1,462 条，发布 744 条，其中 307 条标为推荐。
Published:
What makes a person exceptional? How can we create exceptional things? What separates exceptional work from mediocre work?
Published:
什么样的人是杰出的？我们怎么样才能造出杰出的东西？杰出的东西和平庸的东西有什么区别？
Published:
A high school student can now submit to Science. The student did not pick the question.
Published:
高中生也能投 Science 了，但题目不是高中生选的。
Published:
Day 103 of Benchmark Radar. The scan fetched 1,627 items, published 831, and marked 444 as recommended.
Published:
Benchmark Radar 第 103 天。基准和软件一样，会积累维护债务。本轮抓取 1,627 条，发布 831 条，其中 444 条标为推荐。
Published:
Day 102 of Benchmark Radar. The scan fetched 1,638 items, published 785, and marked 354 as recommended.
Published:
Benchmark Radar 第 102 天。没有证据的验证器，只是在发表意见。本轮抓取 1,638 条，发布 785 条，其中 354 条标为推荐。
Published:
Day 101 of Benchmark Radar. The scan fetched 1,686 items, published 844, and marked 354 as recommended.
Published:
Benchmark Radar 第 101 天。排行榜排的可能是评测框架。本轮抓取 1,686 条，发布 844 条，其中 354 条标为推荐。
Published:
I found something embarrassing about myself. On many nights I cannot sleep while my AI quota still has room left.
Published:
我最近发现一件很丢人的事，AI 额度还没用完，我就睡不着觉。
Published:
Day 100 of Benchmark Radar. The scan fetched 1,225 items, published 604, and marked 327 as recommended.
Published:
Benchmark Radar 第 100 天。长不等于深，轮数测的是模型不是任务。本轮抓取 1,225 条，发布 604 条，其中 327 条标为推荐。
Published:
Day 99 of Benchmark Radar. The scan fetched 1,039 items, published 414, and marked 193 as recommended.
Published:
Benchmark Radar 第 99 天。测试时攻击占八成，部署时只占百分之二。本轮抓取 1,039 条，发布 414 条，其中 193 条标为推荐。
Published:
Day 98 of Benchmark Radar. The scan fetched 950 items, published 357, and marked 150 as recommended.
Published:
Benchmark Radar 第 98 天。知道自己在被测的智能体，是另一个智能体。本轮抓取 950 条，发布 357 条，其中 150 条标为推荐。
Published:
Day 97 of Benchmark Radar. The scan fetched 1,202 items, published 517, and marked 238 as recommended.
Published:
Benchmark Radar 第 97 天。每条回复都没问题，整场对话却可以出事。本轮抓取 1,202 条，发布 517 条，其中 238 条标为推荐。
Published:
Day 96 of Benchmark Radar. The scan fetched 1,461 items, published 676, and marked 332 as recommended.
Published:
Benchmark Radar 第 96 天。先检查技能，再让它运行。本轮抓取 1,461 条，发布 676 条，其中 332 条标为推荐。
Published:
Day 95 of Benchmark Radar. The scan fetched 1,308 items, published 568, and marked 250 as recommended.
Published:
Benchmark Radar 第 95 天。会回答空间问题，不等于会在空间里行动。本轮抓取 1,308 条，发布 568 条，其中 250 条标为推荐。
Published:
Day 94 of Benchmark Radar. The scan fetched 1,283 items, published 573, and marked 238 as recommended.
Published:
Benchmark Radar 第 94 天。正确答案，有时就是根本不存在方案。本轮抓取 1,283 条，发布 573 条，其中 238 条标为推荐。
Published:
Day 93 of Benchmark Radar. The scan fetched 1,280 items, published 643, and marked 320 as recommended.
Published:
Benchmark Radar 第 93 天。评判智能体实际改了什么。本轮抓取 1,280 条，发布 643 条，其中 320 条标为推荐。
Published:
Day 92 of Benchmark Radar. The scan fetched 1,048 items, published 464, and marked 211 as recommended.
Published:
Benchmark Radar 第 92 天。经典文档基准的高分，一部分来自错标签和重复数据。本轮抓取 1,048 条，发布 464 条，其中 211 条标为推荐。
Published:
Day 91 of Benchmark Radar. The scan fetched 1,310 items, published 466, and marked 163 as recommended.
Published:
Benchmark Radar 第 91 天。静态图表分数，看不见图表藏起来的东西。本轮抓取 1,310 条，发布 466 条，其中 163 条标为推荐。
Published:
Day 90 of Benchmark Radar. The scan fetched 1,231 items, published 535, and marked 242 as recommended.
Published:
Benchmark Radar 第 90 天。个人记忆需要权限检查，不只是记得准。本轮抓取 1,231 条，发布 535 条，其中 242 条标为推荐。
Published:
Day 89 of Benchmark Radar. The scan fetched 1,332 items, published 630, and marked 289 as recommended.
Published:
Benchmark Radar 第 89 天。工具说真话，也可能漏掉关键证据。本轮抓取 1,332 条，发布 630 条，其中 289 条标为推荐。
Published:
Day 88 of Benchmark Radar. The scan fetched 1,613 items, published 814, and marked 319 as recommended.
Published:
Benchmark Radar 第 88 天。先冻结策略，再测留出期。本轮抓取 1,613 条，发布 814 条，其中 319 条标为推荐。
Published:
Day 87 of Benchmark Radar. The scan fetched 1,692 items, published 884, and marked 391 as recommended.
Published:
Benchmark Radar 第 87 天。任务成功仍需检查授权。本轮抓取 1,692 条，发布 884 条，其中 391 条标为推荐。
Published:
Day 86 of Benchmark Radar. The scan fetched 1,430 items, published 755, and marked 389 as recommended.
Published:
Benchmark Radar 第 86 天。客服智能体还要检查操作过程。本轮抓取 1,430 条，发布 755 条，其中 389 条标为推荐。
Published:
Day 85 of Benchmark Radar. The scan fetched 1,355 items, published 602, and marked 281 as recommended.
Published:
Benchmark Radar 第 85 天。逐条判断证据是否改变。本轮抓取 1,355 条，发布 602 条，其中 281 条标为推荐。
Published:
Day 84 of Benchmark Radar. The scan fetched 1,233 items, published 488, and marked 165 as recommended.
Published:
Benchmark Radar 第 84 天。每个正确答案花多少钱。本轮抓取 1,233 条，发布 488 条，其中 165 条标为推荐。
Published:
Day 83 of Benchmark Radar. The scan fetched 1,562 items, published 749, and marked 246 as recommended.
Published:
Benchmark Radar 第 83 天。安全分数也取决于检测器版本。本轮抓取 1,562 条，发布 749 条，其中 246 条标为推荐。
Published:
Day 82 of Benchmark Radar. The scan fetched 1,672 items, published 846, and marked 346 as recommended.
Published:
Benchmark Radar 第 82 天。输出得分无法证明运行时完整。本轮抓取 1,672 条，发布 846 条，其中 346 条标为推荐。
Published:
Day 81 of Benchmark Radar. The scan fetched 1,711 items, published 899, and marked 425 as recommended.
Published:
Benchmark Radar 第 81 天。去重后仍要按患者划分。本轮抓取 1,711 条，发布 899 条，其中 425 条标为推荐。
Published:
Day 80 of Benchmark Radar. The scan fetched 1,506 items, published 773, and marked 410 as recommended.
Published:
Benchmark Radar 第 80 天。用实验预算评估科研智能体。本轮抓取 1,506 条，发布 773 条，其中 410 条标为推荐。
Published:
Day 79 of Benchmark Radar. The scan fetched 1,476 items, published 624, and marked 323 as recommended.
Published:
Benchmark Radar 第 79 天。标准答案能否从输入推出。本轮抓取 1,476 条，发布 624 条，其中 323 条标为推荐。
Published:
Day 78 of Benchmark Radar. The scan fetched 997 items, published 449, and marked 223 as recommended.
Published:
Benchmark Radar 第 78 天。注入传到了，不等于执行了。本轮抓取 997 条，发布 449 条，其中 223 条标为推荐。
Published:
Day 77 of Benchmark Radar. The scan fetched 1,263 items, published 477, and marked 164 as recommended.
Published:
Benchmark Radar 第 77 天。记忆能否跟上新要求。本轮抓取 1,263 条，发布 477 条，其中 164 条标为推荐。
Published:
Day 76 of Benchmark Radar. The scan fetched 1,316 items, published 547, and marked 234 as recommended.
Published:
Benchmark Radar 第 76 天。测试集要留出完整录制。本轮抓取 1,316 条，发布 547 条，其中 234 条标为推荐。
Published:
Day 75 of Benchmark Radar. The scan fetched 1,338 items, published 616, and marked 281 as recommended.
Published:
Benchmark Radar 第 75 天。分数变了，先查评分器。本轮抓取 1,338 条，发布 616 条，其中 281 条标为推荐。
Published:
Day 74 of Benchmark Radar. The scan fetched 1,326 items, published 604, and marked 274 as recommended.
Published:
Benchmark Radar 第 74 天。安全监控需要完整日志。本轮抓取 1,326 条，发布 604 条，其中 274 条标为推荐。
Published:
Day 73 of Benchmark Radar. The scan fetched 1,657 items, published 833, and marked 376 as recommended.
Published:
Benchmark Radar 第 73 天。不读消息，怎样审计智能体。本轮抓取 1,657 条，发布 833 条，其中 376 条标为推荐。
Published:
Day 72 of Benchmark Radar. The daily snapshot archive has no record for October 6, 2026, so there are no verifiable fetched, published, or recommended counts, or research highlights, for this date.
Published:
Benchmark Radar 第 72 天。每日快照归档中缺少 2026 年 10 月 6 日的记录，因此当天没有可核对的抓取量、发布量、推荐量或研究摘要。
Published:
Day seventy-one of Benchmark Radar. The daily scan published 560 records and recommended 273, led by “Passing the Test You Trained On,” a study of 15 prompt-injection detectors. Scoreboard: 279 stars, 45 forks.
Published:
Benchmark Radar 的第七十一天。每日扫描发布 560 条记录，推荐 273 条，领头的是 《Passing the Test You Trained On》：一项针对 15 个提示注入检测器的研究。记分牌：279 颗星，45 个 fork。
Published:
Day seventy of Benchmark Radar. The daily scan published 484 records and recommended 127, led by the 2026 Cryo-EM Benchmark Dataset, 12 protein structures deposited after every disclosed test set for MICA. Scoreboard: 278 stars, 45 forks.
Published:
Benchmark Radar 的第七十天。每日扫描发布 484 条记录，推荐 127 条，领头的是 2026 Cryo-EM Benchmark Dataset：12 个蛋白结构，沉积时间晚于 MICA 披露的所有测试集。记分牌：278 颗星，45 个 fork。
Published:
Day sixty-nine of Benchmark Radar. The daily scan published 577 records and recommended 174, led by an updated retinal benchmark audit that publishes image-level manifests and per-image predictions. Scoreboard: 275 stars, 45 forks.
Published:
Benchmark Radar 的第六十九天。每日扫描发布 577 条记录，推荐 174 条，领头的是 一项更新后的视网膜基准审计：它公开逐图像的清单和逐图像的预测。记分牌：275 颗星，45 个 fork。
Published:
Day sixty-eight of Benchmark Radar. The daily scan published 668 records and recommended 244, led by Video-Index, a meta-benchmark that audits whether video tests require video understanding. Scoreboard: 272 stars, 45 forks.
Published:
Benchmark Radar 的第六十八天。每日扫描发布 668 条记录，推荐 244 条，领头的是 Video-Index：这个元基准审计视频测试是否真的需要理解视频。记分牌：272 颗星，45 个 fork。
Published:
Day sixty-seven of Benchmark Radar. The daily scan published 688 records and recommended 289, led by “Dating the Model,” a study showing that a hidden current-date line changes measured performance. Scoreboard: 268 stars, 45 forks.
Published:
Benchmark Radar 的第六十七天。每日扫描发布 688 条记录，推荐 289 条，领头的是 《Dating the Model》：这项研究表明，一行隐藏的当前日期会改变测得的表现。记分牌：268 颗星，45 个 fork。
Published:
Day sixty-six of Benchmark Radar. The daily scan published 656 records and recommended 186, led by UserProxyBench, which tests whether the language model playing the user follows its private instructions. Scoreboard: 268 stars, 43 forks.
Published:
Benchmark Radar 的第六十六天。每日扫描发布 656 条记录，推荐 186 条，领头的是 UserProxyBench：它检验扮演用户的语言模型是否遵守它的私有指令。记分牌：268 颗星，43 个 fork。
Published:
Day sixty-five of Benchmark Radar. The daily scan published 758 records and recommended 381, led by TraceDance, which turns real agent deployment traces into benchmarks for undesirable behavior. Scoreboard: 266 stars, 43 forks.
Published:
Benchmark Radar 的第六十五天。每日扫描发布 758 条记录，推荐 381 条，领头的是 TraceDance：它把真实的智能体部署轨迹变成针对不良行为的基准。记分牌：266 颗星，43 个 fork。
Published:
Day sixty-four of Benchmark Radar. The daily scan published 505 records and recommended 233, led by the Finance Agents Benchmark, which packages financial due diligence as an end-to-end agent test. Scoreboard: 264 stars, 42 forks.
Published:
Benchmark Radar 的第六十四天。每日扫描发布 505 条记录，推荐 233 条，领头的是 Finance Agents Benchmark：它把金融尽职调查做成端到端的智能体测试。记分牌：264 颗星，42 个 fork。
Published:
Day sixty-three of Benchmark Radar. The daily scan published 314 records and recommended 106, led by the Kalyvox Voice Benchmark 2026, which scores one voice agent across 240 controlled calls. Scoreboard: 257 stars, 41 forks.
Published:
Benchmark Radar 的第六十三天。每日扫描发布 314 条记录，推荐 106 条，领头的是 Kalyvox Voice Benchmark 2026：它用 240 通受控通话给一个语音智能体打分。记分牌：257 颗星，41 个 fork。
Published:
Day sixty-two of Benchmark Radar. The daily scan published 540 records and recommended 171, led by repo2bench, which turns Python repositories into coding-agent tasks. Scoreboard: 255 stars, 41 forks.
Published:
Benchmark Radar 的第六十二天。每日扫描发布 540 条记录，推荐 171 条，领头的是 repo2bench：它把 Python 仓库转成编程智能体的任务。记分牌：255 颗星，41 个 fork。
Published:
Day sixty-one of Benchmark Radar. The daily scan published 450 records and recommended 198, led by PartHackBench, which tests whether partial-credit scorers reward misleading progress in long-running tool agents. Scoreboard: 252 stars, 41 forks.
Published:
Benchmark Radar 的第六十一天。每日扫描发布 450 条记录，推荐 198 条，领头的是 PartHackBench：它检验部分得分机制是否会奖励长时间运行的工具智能体的误导性进展。记分牌：252 颗星，41 个 fork。
Published:
Day sixty of Benchmark Radar. Benchmark Radar reaches 250 stars on day sixty. The daily scan published 471 records and recommended 199, led by Uncheatable Eval, which scores models on text published after they were trained. Scoreboard: 250 stars, 41 forks.
Published:
Benchmark Radar 的第六十天。第六十天，Benchmark Radar 达到 250 颗星。每日扫描发布 471 条记录，推荐 199 条，领头的是 Uncheatable Eval：它用模型训练之后才发表的文本来打分。记分牌：250 颗星，41 个 fork。
Published:
Have you ever run into this while using AI?
Published:
你在用 AI 的时候，有没有遇到过这种情况？
Published:
Day fifty-nine of Benchmark Radar. The daily scan published 510 records and recommended 211, led by a molecular audit that finds hidden duplicates across train and test splits and a benchmark for coding agents whose patches break when merged. Scoreboard: 249 stars, 41 forks.
Published:
Benchmark Radar 的第五十九天。每日扫描发布 510 条记录，推荐 211 条。领头的是一项发现训练集与测试集之间隐藏重复的分子数据审计，以及一个检验编程智能体补丁合并后是否冲突的基准。记分牌：249 颗星，41 个 fork。
Published:
Day fifty-eight of Benchmark Radar. Two fixes clear yesterday’s stop without loosening the check that caused it: Zenodo file batches collapse to one finding, and copied Hugging Face template text no longer passes as a description. Scoreboard: 246 stars, 41 forks.
Published:
Benchmark Radar 的第五十八天。两个修复解除了昨天的停机，却没有放松触发停机的检查：Zenodo 的分文件批量存档合并为一条发现，Hugging Face 模板的复制文本不再冒充描述。记分牌：246 颗星，41 个 fork。
Published:
Day fifty-seven of Benchmark Radar. The daily run failed on purpose. A safety check found 16 records sharing one description and refused to publish them. Scoreboard: 238 stars, 41 forks.
Published:
Benchmark Radar 的第五十七天。每日运行按设计失败了。一项安全检查发现 16 条记录共用同一段描述，拒绝发布。记分牌：238 颗星，41 个 fork。
Published:
Day fifty-six of Benchmark Radar. Contributor winklemad lands two fixes: one makes record-holder credit agree with the record chart, the other keeps every benchmark’s page address unique. Scoreboard: 236 stars, 41 forks.
Published:
Benchmark Radar 的第五十六天。贡献者 winklemad 合并两个修复：一个让纪录保持者的归属与纪录图一致，另一个保证每个基准的页面地址唯一。记分牌：236 颗星，41 个 fork。
Published:
Day fifty-five of Benchmark Radar. The contributor guide gains one rule: start every branch from an up-to-date main. The WeChat group QR code is refreshed, and the daily scan ran on schedule. Scoreboard: 234 stars, 40 forks.
Published:
Benchmark Radar 的第五十五天。贡献者指南新增一条规则：每个分支都从最新的 main 开始。微信群二维码更新，每日扫描按时运行。记分牌：234 颗星，40 个 fork。
Published:
Day fifty-four of Benchmark Radar. The daily scan published 618 records and recommended 224, including a 1.6-million-mask ultrasound release and OverclaimBench, a test of whether coding agents falsely report finished work. Scoreboard: 233 stars, 38 forks.
Published:
Benchmark Radar 的第五十四天。每日扫描发布 618 条记录，推荐 224 条，其中包括一个含 160 万张分割掩码的超声数据发布，以及检验编程智能体是否谎报完工的 OverclaimBench。记分牌：233 颗星，38 个 fork。
Published:
Day fifty-three of Benchmark Radar. The daily scan published 666 records and recommended 261 of them, led by a corrected word-sense benchmark, a stablecoin stress arena, and a ride-hailing agent test. Scoreboard: 229 stars, 36 forks.
Published:
Benchmark Radar 的第五十三天。每日扫描发布 666 条记录，推荐其中 261 条。领头的是一个修正过的词义基准、一个稳定币压力测试场，以及一个网约车智能体测试。记分牌：229 颗星，36 个 fork。
Published:
Day fifty-two of Benchmark Radar. The dataset ships to Hugging Face with automated sync, and seventeen project-only words leave the interface for plain language. Scoreboard: 226 stars, 35 forks.
Published:
Benchmark Radar 的第五十二天。数据集随自动同步上线 Hugging Face，十七处项目黑话退出界面，换成直白语言。记分牌：226 颗星，35 个 fork。
Published:
Day fifty-one of Benchmark Radar. A quiet day with loud groundwork: the daily snapshot recorded while the Hugging Face dataset export goes through validation. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十一天。安静的一天，底下的功夫并不轻：每日快照照常入库，Hugging Face 数据集导出走完验证。记分牌：185 颗星，30 个 fork。
Published:
A laser lightsail accelerating to one percent of light speed turns about one percent of the beam into motion. The other ninety-nine percent leaves as light and does no work.
Published:
激光光帆加速到光速的百分之一，光束里大约只有百分之一变成了运动。剩下的百分之九十九以光的形式离开，不做功。
Published:
Day fifty of Benchmark Radar. The paper links point at arXiv everywhere, the README asks readers to upvote on Hugging Face, and a dated laurel badge marks the #1 Paper of the Day on the dashboard. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十天。论文链接全项目指向 arXiv，README 请读者去 Hugging Face 投票，仪表盘挂起带日期的每日第一论文桂冠。记分牌：185 颗星，30 个 fork。
Published:
The trendiest sentence right now is “this tastes too much like AI, we need to de-AI it.”
Published:
现在最时髦的一句话，是“这东西 AI 味太重，得 de-AI”。
Published:
Day forty-nine of Benchmark Radar. The README starts naming its readers: a researchers-by-institution strip with official logos, built by a script, plus a fresh WeChat group code for daily discussion. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十九天。README 开始点名读者：按机构排列的研究者横幅，官方 logo 由脚本生成；微信群二维码更新，每日基准动态的中文讨论室继续开放。记分牌：185 颗星，30 个 fork。
Published:
Day forty-eight of Benchmark Radar. The dashboard gets a reader-first pass: the homepage banner keeps its words but drops its pixels, benchmark pages answer what readers actually search for, and an About page tells the project’s story on the site. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十八天。仪表盘迎来读者优先改造：首页横幅保留文字、去掉像素占用，基准页围绕读者真实搜索重写，About 页把项目故事讲在站内。记分牌：185 颗星，30 个 fork。
Published:
Day forty-seven of Benchmark Radar. With the arXiv paper live, the citation surfaces get a reader-first refresh: the Cite card collapses APA and BibTeX blocks by default, CITATION.md lands in the repo root, and a Cite link joins the top navigation bar. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十七天。arXiv 论文上线后，引用位置全面转为读者优先：Cite 卡片默认折叠 APA 与 BibTeX，仓库根目录落地 CITATION.md，顶部导航新增 Cite 入口。记分牌：185 颗星，30 个 fork。
Published:
Day forty-six of Benchmark Radar. The technical paper is now on arXiv as arXiv:2609.11115, the v0.11.0 release is published with frozen data and verified hashes, and citation surfaces across the project update to the live arXiv entry. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十六天。技术论文正式登上 arXiv，编号 arXiv:2609.11115；v0.11.0 发布携带冻结数据与经验证哈希；全项目引用位置更新指向在线 arXiv 条目。记分牌：185 颗星，30 个 fork。
Published:
Day forty-five of Benchmark Radar. The paper enters its final preparation phase: content relicensed to CC BY-NC-SA 4.0, CI checks validate frozen paper data, and report tests align with the manuscript. Scoreboard: 180 stars, 32 forks.
Published:
Benchmark Radar 的第四十五天。论文进入最后准备阶段：编辑内容改授 CC BY-NC-SA 4.0，CI 核验冻结论文数据，报告测试与手稿对齐。记分牌：180 颗星，32 个 fork。
Published:
The Navier-Stokes fight is not proof that anyone stole anything. It is proof that the person whose work was at stake had no way to check.
Published:
Navier-Stokes 之争并没有证明谁偷了谁的东西。它证明的是：那个把研究押在里面的人，根本无从查证。
Published:
Day forty-four of Benchmark Radar. The paper submodule advances through the citation revision and the source-count revision to the rebuilt PDF. The arXiv submission package assembles. Scoreboard: 175 stars, 31 forks.
Published:
Benchmark Radar 的第四十四天。论文子模块经引文修订、来源计数修订推进到重建 PDF。arXiv 投稿包成型。记分牌：175 颗星，31 个 fork。
Published:
Day forty-three of Benchmark Radar. The discovery cutoff for the technical paper locks at 2026-09-07. Source counts, figure data, and the full-catalog audit pipeline freeze for the v0.11.0 release and arXiv submission. Scoreboard: 170 stars, 30 forks.
Published:
Benchmark Radar 的第四十三天。技术论文的发现截止锁定在 2026-09-07。来源计数、图表数据、全目录审计管道全部冻结，服务 v0.11.0 发布与 arXiv 投稿。记分牌：170 颗星，30 个 fork。
Published:
Day forty-two of Benchmark Radar. The ranking engine for recent releases moves from backend to the dashboard, and the daily radar snapshot captures the first full week of ranked cohorts. Scoreboard: 165 stars, 29 forks.
Published:
Benchmark Radar 的第四十二天。新鲜发布的排行榜引擎从后端走到仪表盘，每日雷达快照记下第一个完整的排序周期。记分牌：165 颗星，29 个 fork。
Published:
The dark night gave me dark eyes, yet I use them to seek the light.
Published:
黑夜给了我黑色的眼睛，我却用它寻找光明。
Published:
Day forty-one of Benchmark Radar. The ranking engine for recent releases shipped its first phase with the data contract written first, and the report audit separated raw best scores from scores measured under the same protocol. Scoreboard: 160 stars, 28 forks.
Published:
Benchmark Radar 的第四十一天。新鲜发布的排行榜引擎第一阶段上线，先把数据契约写在前头；报告审计也把原始最高分和同一协议下测出的分数分开。记分牌：160 颗星，28 个 fork。
Published:
Day forty of Benchmark Radar. The radar started watching the network where most new benchmark papers get their DOI, brought a benchmark back by searching its exact name, and started posting where it was not posting before. Scoreboard: 147 stars, 27 forks.
Published:
Benchmark Radar 的第四十天。雷达开始盯着大多数新基准论文拿到 DOI 的那个网络，靠搜索一个基准的精确名字把它找了回来，并开始向以前没发过的地方发布。记分牌：147 颗星，27 个 fork。
Published:
Day thirty-nine of Benchmark Radar. The daily brief now has a landing page, a full archive, and a feed of its own, and the dashboard navigation got short enough to scan. Scoreboard: 144 stars, 27 forks.
Published:
Benchmark Radar 的第三十九天。每日简报现在有了自己的落地页、完整存档和订阅源，仪表盘导航也短到一眼能扫完。记分牌：144 颗星，27 个 fork。
Published:
Day thirty-eight of Benchmark Radar. The daily brief, the thing the project already wrote every day and never gave a page, now has a blog and a feed. Scoreboard: 137 stars, 26 forks.
Published:
Benchmark Radar 的第三十八天。每日简报，这个项目每天都写、却从没给它一个页面的东西，现在有了博客和订阅源。记分牌：137 颗星，26 个 fork。
Published:
Day thirty-seven of Benchmark Radar. Anyone writing a paper can now copy a citation without leaving the dashboard, the offline setup moved onto the site, and search looks through every collected day by default. Scoreboard: 130 stars, 25 forks.
Published:
Benchmark Radar 的第三十七天。写论文的人现在不用离开仪表盘就能复制引用，离线安装搬到了网站上，搜索默认会翻遍每一个采集过的日子。记分牌：130 颗星，25 个 fork。
Published:
Day thirty-six of Benchmark Radar. The biggest model card yet landed with 37 scores across 34 benchmarks, and search stopped guessing which result answered the question. Scoreboard: 128 stars, 24 forks.
Published:
Benchmark Radar 的第三十六天。迄今最大的一张模型卡落地了，34 个基准上排了 37 个分数；搜索也不再靠猜来回答哪个结果是对的。记分牌：128 颗星，24 个 fork。
Published:
Day thirty-five of Benchmark Radar. A week-old copycat outranked the real project for its own name because search engines could see only four pages here. Today every one of the 1,173 benchmarks the radar tracks has a page of its own. Scoreboard: 123 stars, 23 forks.
Published:
Benchmark Radar 的第三十五天。一个上线才一周的仿站，靠自家名字把真项目压在了搜索结果下面，因为搜索引擎在这里只能看到四个页面。今天，雷达追踪的 1173 个基准，每一个都有了自己独立的页面。记分牌：123 颗星，23 个 fork。
Published:
Day thirty-four of Benchmark Radar. The radar now runs from a command line on your own computer, offline, no dashboard in between, and a coding agent can answer benchmark questions straight from the data. Scoreboard: 121 stars, 23 forks.
Published:
Benchmark Radar 的第三十四天。雷达现在可以从你自己电脑的命令行运行，离线可用，中间不隔仪表盘，编码智能体可以直接从数据里回答基准测试的问题。记分牌：121 颗星，23 个 fork。
Published:
Day thirty-three of Benchmark Radar. The radar learned to find benchmark releases that keyword search would never catch, and a new frontier model card with 14 benchmarks joined the registry. Scoreboard: 115 stars, 21 forks.
Published:
Benchmark Radar 的第三十三天。雷达学会了找到关键词搜索永远抓不到的基准发布，注册表里也进了一张带 14 个基准的新前沿模型卡片。记分牌：115 颗星，21 个 fork。
Published:
Day thirty-two of Benchmark Radar. We fixed data errors that failed silently, taught the ranking to count adoption for releases that ship no binary files, and rewrote the daily briefing so it reads like a person wrote it. Scoreboard: 113 stars, 21 forks.
Published:
Benchmark Radar 的第三十二天。我们修掉了静默失败的数据错误，让排名在发布不附带任何二进制文件时也能统计采纳，还把每日简报改写成像人写的。记分牌：113 颗星，21 个 fork。
Published:
Day thirty-one of Benchmark Radar. Every release card now carries a one-line summary instead of a bare version tag, filing an issue takes at most two required fields, and both READMEs got short enough to read. Scoreboard: 107 stars, 20 forks.
Published:
Benchmark Radar 的第三十一天。每张发布卡片现在都带一行摘要，而不是一个光秃秃的版本号；提交 issue 最多填两个必填项；两份 README 也精简到读得下去。记分牌：107 颗星，20 个 fork。
Published:
There are only 26 single-letter .ai domains. I went through all of them, one by one, to see who actually builds something and who just squats. The answer: about six run real products, one belongs to Elon Musk, one belongs to Google, one apparently belongs to Apple without a website, and the rest are for-sale pages asking between 1.5 and 500 million dollars. Here is the full alphabet.
Published:
单字母 .ai 域名只有 26 个。我把它们挨个访问了一遍，看谁在真做事，谁在蹲着等涨价。答案：大约六个在跑真产品，一个属于马斯克，一个属于谷歌，一个据说属于苹果但没有网站，剩下的是开价 150 万到 5 亿美元的出售页。以下是完整字母表。
Published:
Day thirty of Benchmark Radar. Thirty posts in, this log has become a live broadcast of one open-source project trying to reach 1,000 stars, and today the scoreboard reads 86.
Published:
Benchmark Radar 的第三十天。写到今天，这份日志正式变成一场公开直播：一个开源项目如何从零走向 1000 颗星。先报分数：今天记分牌上写着 86。
Published:
Day twenty-nine of Benchmark Radar. We ranked new releases before routine updates, put a 0 to 1 score on a 0 to 100 scale without changing its value, and made the active tab stay active.
Published:
新东西该在第一页，而不是藏在第三页。第二十九天我们做了三件事：让新发布排在例行更新之前，把 0 到 1 的分数按 0 到 100 来显示而不断点，让当前栏目始终看起来是选中的。先说几个词：发布是全新冒出来的基准；更新是对已有条目的改动，例如版本升级或新增分数；0 到 1 与 0 到 100 可以是同一数值的两种刻度。
Published:
Day twenty-eight of Benchmark Radar. We removed a second copy of the data, made the page say one thing to crawlers, and gave the work a citable name.
Published:
别给同一份真相留两个副本，否则它们迟早对不上。第二十八天我们做了三件事：把可生成的数据移出版本库，让页面只对爬虫说一个标题，给作品一个可引用的名字。先说几个词：真相来源是唯一可信的那份源文件，其余都由它生成；h1 是页面的主标题，爬虫期望整页只有一个；i18n 是国际化，让界面按语言显示；引用是你写论文时粘贴的那条参考文献。
Published:
Day twenty-seven of Benchmark Radar. We made the Today list load a page at a time, cut a title in half, and taught 21 benchmarks to show their real names.
Published:
一天 136 条还要一次全画出来，只会让首屏变慢。第二十七天我们做了三件事：让今日列表一页页加载，把标题砍半，让 21 个基准显示真名。先说几个词：分页是一长串内容分多页看完；首包是页面为了快而先加载的小数据包；SEO 是让搜索引擎看懂并收录网站的做法。
Published:
Day twenty-six of Benchmark Radar. We put the ranking where it belongs, made the Back button work, and replaced an invented chart with one we can draw.
Published:
一个以排名命名的页面，却把排名藏在第六屏，这不叫排名页。第二十六天我们做了三件事：让排名站到最前，让返回键能用，把一条编出来的图换成一条能画的线。先说几个词：排名就是按多少排序的榜单；返回键是浏览器左上角让你回到上一页的箭头；前沿线是一条阶梯线，连起截至每个日期为止的最好分数；过期横幅是数据过时时顶部的提示条。
Published:
Day twenty-five of Benchmark Radar. We stopped the radar from ranking itself, put scores on the dates they belong to, and let a benchmark with no score still speak for itself.
Published:
一个会把自己排进榜单的雷达，不配叫雷达。第二十五天我们做了三件事：让排名不再跟自己玩博弈，让每天抓来的分数落在它该在的发布日期上，再让那些还没人评分的基准自己说出身份。先说一个词：排行榜（leaderboard）就是一张公开榜单，谁分数高谁排前面。
Published:
Day twenty-four of Benchmark Radar. We put the results first on the homepage, and we stopped blaming you when a page came up empty.
Published:
今天我们把首页重做了一遍，让它先把结果摆出来，而不是先甩给读者一个空页面。我们还让那些「爬到了但没结果」的来源变得看得见了。先解释两个词：PR 是别人提议改的一段代码，issue 是一个待办或 bug。
Published:
Day twenty-three of Benchmark Radar. We started pulling in other people’s benchmark lists, and we made the leaderboard searchable.
Published:
雷达吸收了外部目录。第二十三天构建了外部目录系统，并在排行榜中新增了基准测试搜索功能。
Published:
Day twenty-two of Benchmark Radar. We locked down data integrity and opened up the community.
Published:
过时运行不再能污染快照数据，微信群正式公开。第二十二天强化了数据完整性并扩展了社区渠道。
Published:
Hi, Koutian here. Day twenty-one was a fix day that touched many parts at once.
Published:
OpenReview 完成了认证接入，宽屏布局得以扩展，手机端的显示问题也修复了。第二十一天是横跨多个模块的修复日。
Published:
Hi, Koutian here. Day twenty taught the radar to speak Chinese and to summarize itself.
Published:
雷达开始说中文了。第二十天交付了完整的 zh 国际化词典、洞察块和完整的社交清单。
Published:
Hi, Koutian here. Day nineteen pushed the radar into maps and satellites.
Published:
雷达扩展到了地理空间和卫星 AI 领域。第十九天为饱和度图添加了领域特定信号和厂商 Logo。
Published:
Hi, Koutian here. Day eighteen was a quiet day, and that was the point.
Published:
你好，我是 Koutian。第十八天，管道自己跑完了，没要任何人插手。这一天，恰恰证明了它稳。
Published:
Hi, Koutian here. Day seventeen was a cleanup day. We fixed what an internal audit found and made the site look finished.
Published:
你好，我是 Koutian。第十七天是加固与美化日。我们没加什么大功能，而是把之前埋的雷排掉，再把界面收拾干净。审计发现的问题都处理完了。
Published:
Some benchmarks were counted wrong, and the dashboards were overwhelming. Day sixteen fixed both.
Published:
你好，我是 Koutian。第十六天，我们修了一个会悄悄骗人的数，还让界面不再一打开就吓到人。
Published:
The radar started posting to social media. Day fifteen built that pipeline and retired the daily GitHub Issue.
Published:
The daily pipeline went from fragile to reliable. Day fourteen added more feeds, hardened the briefing, and turned on Q&A in production.
Published:
The radar now backs up every claim with a link. Day thirteen added evidence grounding, a daily Q&A on the dashboard, and a stack of small fixes.
Published:
The radar now has an RSS feed. Day twelve was a small day, one feature, but it changes how you follow the project.
Published:
KW-Bench got a capability rubric. The community got launch posts. Hi, Koutian here. Day eleven ran on two tracks at once.
Published:
The radar started talking. Hi, Koutian here. Day ten added AI-written briefings, GPT-powered insights, and launch copy for Chinese social platforms.
Published:
雷达开始会自己说话了。第十天我们加了 AI 生成的每日简报、GPT 驱动的洞察，还有面向中文社交平台的发布文案。
Published:
The leaderboard became a research workbench. Hi, Koutian here. Day nine added the adoption frontier, score progression, and saturation charts.
Published:
排行榜以前只是一张排名表。第九天，我们把它改成了能回答问题的研究工作台，加了采纳前沿、分数进展和饱和度可视化。
Published:
A quiet day after the registry explosion. Hi, Koutian here. Day eight settled the pipeline with two precise fixes.
Published:
注册表大爆炸之后，安静的一天。第八天，我们用两次精准修复稳住了流水线。
Published:
Day seven was the biggest single day. Hi, Koutian here. We launched a new leaderboard, grew the registry to the 2026 frontier, and bumped the version to 0.3.0.
Published:
最大的一天。模型卡采纳排行榜上线，注册表扩到了 2026 年前沿，版本号升到 0.3.0。
Published:
The radar now watches Hacker News. Day six added community attention signals and hardened the daily schedule.
Published:
雷达开始听 Hacker News 的动静了。第六天，我们加了社区注意力信号，还加固了每天定时运行的可靠性。
Published:
The trend chart got a real hover card. The landscape report got published. And the agentic benchmark count jumped from 3 to 78.
Published:
趋势图现在有了一个真正的悬浮卡片。全景分析报告发布了。agentic 基准的数量从 3 跳到了 78。
Published:
The radar learned to remember. Day four added historical backfill, staleness detection, and an agentic taxonomy category.
Published:
雷达学会了记忆。第四天，我们加了历史数据回填、过期检测，还有 agentic 这个分类。
Published:
A small day with two precise fixes that made the radar shareable.
Published:
内容不多，但两处精准的修复让雷达变得能分享了。
Published:
AI benchmarks now show up faster than any one researcher can keep up with. So I built a radar that checks for them every day, and shows its work.
Published:
新的 AI 基准冒出来的速度，比任何研究员评估它们的速度都快。所以我建了一个雷达，让「发现新基准」变成一件每天做、过程透明、可以复核的事。先说一个词：基准（benchmark）就是用来考 AI 的一套题，或者一份评测。
Published:
The radar gained memory. Day two added trend maps that span time, plus a fix for benchmarks showing up under many names.
Published:
雷达开始有记忆了。第二天我们建起了累积趋势图，还顺手解决了工件别名的问题。先说一个词：工件（artifact）就是被追踪的每一个具体东西，比如某个基准或数据集。
Published:
Day one of Benchmark Radar. We went from nothing to a working daily dashboard in a single day.
Published:
Benchmark Radar 第一天。我们从零开始，一天之内搭出了一个能跑的累积仪表板。
Published:
The most valuable application of LLMs may not be answering questions, but turning an expert job into a digital work unit that can be executed, reviewed, and audited.
Published:
LLM 最有价值的应用，可能不是回答问题，而是把一项专家工作变成可以执行、复核和审计的数字工作单元。
Published:
The most worth-watching thing in 2030 is not how many people are replaced by AI, but which jobs will be reorganized.
Published:
2030 年最值得观察的，不是有多少人被 AI 取代，而是哪些工作会被重新组织。
Published:
| Original source | Tencent Tech; transcript edited | Gu Lingyu; original editors | Xu Qingyang, Su Yang; 2030 annotations and second edit | Koutian Wu |
Published:
原文来源｜腾讯科技；原文整理｜顾翎羽；原文编辑｜徐青阳、苏扬；2030 注释与二次编辑｜Koutian Wu
Published:
When someone publicly told Steve Jobs that he did not understand technology, Jobs paused, conceded part of the criticism, and asked a more consequential question: what will the customer ultimately gain?
Published:
一个人当众说 Steve Jobs 根本不懂技术，他停了十几秒，没有反击，只问了一件更重要的事，用户最终能得到什么？
Published:
Checking flights today, I suddenly noticed: AI is the airline code for Air India.
Published:
今天看航班时，我突然发现：AI 是 Air India（印度航空）的航空公司代码。
Published:
The thing that almost stranded me on this international trip was not a typhoon, but several unrelated systems deciding at the same time that they did not trust me.
Published:
这次国际旅行真正差点把我留在原地的，不是台风，而是几个互不认识的系统同时决定不相信我。
Published:
Does building a benchmark count as research, or is it “crowdfunding a paper”? Does it create public value or consume public resources? Can it build lasting capabilities, or is it useful only for a startup team trying to show investors its potential?
Published:
Benchmark 到底是在做研究，还是在“众筹一篇论文”？它是在创造公共价值，还是在占用公共资源？它能不能形成长期能力，还是只适合创业团队向投资人展示潜力？
Published:
The most dangerous failure mode of an AI science benchmark is not that it is too hard, it is that it quietly becomes either a coding trick or a guessing game.
Published:
AI 科学 Benchmark 最危险的失败方式，不是它太难，而是它悄悄变成了一道编程技巧题或猜谜题。
Published:
Compute / tokens are the new generation of free eggs. Right now all the big vendors are handing out compute like they’re handing out eggs. When you’re old, you’ll be fighting other old grannies and grandpas for free eggs, this is no longer just a joke, it might be reality.
Published:
算力 / token 是新一代的鸡蛋。现在各大厂商都在发算力，就像在发鸡蛋一样。你老了就要去跟别的老奶奶、老大爷抢发鸡蛋——这已经不是一个笑话，可能就是一个现实。
Published:
What I actually want to do today is pull myself out from under a mountain of repetitive work.
Published:
我今天真正想做的事情，其实是把自己从一座 mountains of 重复性的 work 里面捞出来。
Published:
When you go looking for a seat in the great theater of the AI era, built from PowerPoints, papers, and anxiety, it helps to look at which play each person on stage is performing.
Published:
当你在这个由PPT、Paper和焦虑构成的AI时代大剧院里找座位时，不如先看看台上的人都在演哪一出戏。
Published:
I spent the last couple of days checking a tempting claim: that Microsoft Flight Simulator is secretly the foundation for a broader Earth digital twin story.
Published:
这两天我在核查一个很容易写得很爽、但也很容易写过头的说法：微软模拟飞行（Microsoft Flight Simulator）是不是某种更大地球数字孪生叙事背后的秘密技术底座。
Published:
Anthropic’s most ruthless move isn’t polishing every product to perfection before release, but daring to throw out a research preview first, then rapidly getting stronger in front of everyone.
Published:
Anthropic 最狠的地方，不是每次都把产品打磨到完美才发布，而是敢先把一个 research preview 扔出来，然后当着所有人的面快速变强。
Published:
Prompt Caching Is Not Technical Debt. It’s what the math demands.
Published:
很多人把 prompt caching 看成一个省钱 hack，但我觉得这个判断刚好反了。
Published:
这是我见过最糟糕的创业产品。
Published:
我越来越鄙弃和蔑视「学习」这个词。
Published:
When an AI writing assistant turns a paper’s incoherent rambling into a “well-structured pile of something,” we have to admit: it is basically a cyber laxative.
Published:
当AI写作助手把学术论文里的“狗屁不通”梳理成“结构完整的一大坨”时，我们不得不承认：这简直是一剂赛博开塞露。
Published:
I was in the middle of setting up a second Cloudflare backup, verifying one dataset and configuring another, when the AI agent helping me flagged something unrelated: while listing what had access to my GitHub account, one authorization stood out as unusually broad. That single flag turned into an hour of reviewing GitHub’s installed-apps list, and it was worth every minute.
Published:
我当时正在配置第二份 Cloudflare 备份——核实一份数据、再搭建另一份——帮我干活的 AI Agent 顺手提醒了我一件不相关的事：在列出哪些东西能访问我的 GitHub 账号时，有一项授权的范围明显大得不正常。就是这一句提醒，让我花了一个多小时去仔细过了一遍 GitHub 的已安装应用列表，而这一个小时非常值得。
Published:
I ran a health check on my Claude Code setup this week and found 174 custom skills, 124 of which I had never invoked once. They were not failures. Most of them were scaffolding I built for a model that needed it, and then kept long after the model stopped needing it.
Published:
这周，我给自己的 Claude Code 配置做了一次健康检查，发现里面有 174 个自定义 Skill，其中 124 个我一次都没有调用过。它们并不是失败品。大部分只是我为当时需要额外支撑的模型搭起的脚手架，而模型早已不再需要它们，我却一直把它们留到了现在。
Published:
The last mile of an AI-generated travel plan may be one missing Retry button.
Published:
闲聊历史，美国理科生比商科、法学、医学要给的 stipend 多，因为理科生穷。
Published:
One account opened 58 pull requests in 10 days on benchmark-radar. Five were merged. I put the contribution policy into agent-pr-gate so other maintainers can copy it: before an AI agent starts PR work, it asks its human four questions and waits for the answers.
Published:
一个账号，10 天给 benchmark-radar 提了 58 个 PR，最后合并了 5 个。我把贡献规则整理成了 agent-pr-gate，方便其他维护者复制：AI agent 开始写 PR 前，先问使用它的人四个问题，等人回答后再动手。
Published:
一份由 AI 制定的完美旅行计划，最后可能会死在一个不存在的“重试”按钮上。
Published:
You apply to Anthropic, land an offer, then ask HR whether they’ll sponsor your PERM, and they say “on hold for now” or change the subject, and you probably stand there frozen.
Published:
你投了 Anthropic，拿到了 offer，然后去问 HR 能不能帮你办 PERM，对方说「目前暂停」或者直接转移话题，你当时大概就愣在原地了。
Published:
The trend chart got a real hover card. The landscape report got published. And the agentic benchmark count jumped from 3 to 78.
Published:
趋势图现在有了一个真正的悬浮卡片。全景分析报告发布了。agentic 基准的数量从 3 跳到了 78。
Published:
The radar learned to remember. Day four added historical backfill, staleness detection, and an agentic taxonomy category.
Published:
雷达学会了记忆。第四天，我们加了历史数据回填、过期检测，还有 agentic 这个分类。
Published:
The most dangerous failure mode of an AI science benchmark is not that it is too hard, it is that it quietly becomes either a coding trick or a guessing game.
Published:
AI 科学 Benchmark 最危险的失败方式，不是它太难，而是它悄悄变成了一道编程技巧题或猜谜题。
Published:
My verdict on Air Canada after flying from Singapore to Austin through Vancouver is simple: what a mess.
Published:
这次从新加坡经温哥华飞回奥斯汀，我对 Air Canada 的评价很简单：太拉了。
Published:
Anthropic’s most ruthless move isn’t polishing every product to perfection before release, but daring to throw out a research preview first, then rapidly getting stronger in front of everyone.
Published:
Anthropic 最狠的地方，不是每次都把产品打磨到完美才发布，而是敢先把一个 research preview 扔出来，然后当着所有人的面快速变强。
Published:
You apply to Anthropic, land an offer, then ask HR whether they’ll sponsor your PERM, and they say “on hold for now” or change the subject, and you probably stand there frozen.
Published:
你投了 Anthropic，拿到了 offer，然后去问 HR 能不能帮你办 PERM，对方说「目前暂停」或者直接转移话题，你当时大概就愣在原地了。
Published:
When someone publicly told Steve Jobs that he did not understand technology, Jobs paused, conceded part of the criticism, and asked a more consequential question: what will the customer ultimately gain?
Published:
一个人当众说 Steve Jobs 根本不懂技术，他停了十几秒，没有反击，只问了一件更重要的事，用户最终能得到什么？
Published:
Day forty-three of Benchmark Radar. The discovery cutoff for the technical paper locks at 2026-09-07. Source counts, figure data, and the full-catalog audit pipeline freeze for the v0.11.0 release and arXiv submission. Scoreboard: 170 stars, 30 forks.
Published:
Benchmark Radar 的第四十三天。技术论文的发现截止锁定在 2026-09-07。来源计数、图表数据、全目录审计管道全部冻结，服务 v0.11.0 发布与 arXiv 投稿。记分牌：170 颗星，30 个 fork。
Published:
Day forty-one of Benchmark Radar. The ranking engine for recent releases shipped its first phase with the data contract written first, and the report audit separated raw best scores from scores measured under the same protocol. Scoreboard: 160 stars, 28 forks.
Published:
Benchmark Radar 的第四十一天。新鲜发布的排行榜引擎第一阶段上线，先把数据契约写在前头；报告审计也把原始最高分和同一协议下测出的分数分开。记分牌：160 颗星，28 个 fork。
Published:
Hi, Koutian here. Day seventeen was a cleanup day. We fixed what an internal audit found and made the site look finished.
Published:
你好，我是 Koutian。第十七天是加固与美化日。我们没加什么大功能，而是把之前埋的雷排掉，再把界面收拾干净。审计发现的问题都处理完了。
Published:
The radar now has an RSS feed. Day twelve was a small day, one feature, but it changes how you follow the project.
Published:
At Singapore Changi Airport I saw an ad reading “Asia’s Safest Bank,” and my first question wasn’t how impressive it is, it was, who says so?
Published:
我在新加坡樟宜机场看到一块写着「亚洲最安全银行」的广告，脑子里第一个问题不是它有多厉害，而是，谁说的？
Published:
Does building a benchmark count as research, or is it “crowdfunding a paper”? Does it create public value or consume public resources? Can it build lasting capabilities, or is it useful only for a startup team trying to show investors its potential?
Published:
Benchmark 到底是在做研究，还是在“众筹一篇论文”？它是在创造公共价值，还是在占用公共资源？它能不能形成长期能力，还是只适合创业团队向投资人展示潜力？
Published:
The most dangerous failure mode of an AI science benchmark is not that it is too hard, it is that it quietly becomes either a coding trick or a guessing game.
Published:
AI 科学 Benchmark 最危险的失败方式，不是它太难，而是它悄悄变成了一道编程技巧题或猜谜题。
Published:
Day 110 of Benchmark Radar. The scan fetched 1,548 items, published 676, and marked 352 as recommended.
Published:
Benchmark Radar 第 110 天。你的编程智能体从没见过机器人。本轮抓取 1,548 条，发布 676 条，其中 352 条标为推荐。
Published:
Day 109 of Benchmark Radar. The scan fetched 1,181 items, published 467, and marked 228 as recommended.
Published:
Benchmark Radar 第 109 天。500 项任务里 433 项是占位符。本轮抓取 1,181 条，发布 467 条，其中 228 条标为推荐。
Published:
Day 108 of Benchmark Radar. The scan fetched 1,292 items, published 573, and marked 249 as recommended.
Published:
Benchmark Radar 第 108 天。你的排行榜分不清谁真正赢了。本轮抓取 1,292 条，发布 573 条，其中 249 条标为推荐。
Published:
Day 107 of Benchmark Radar. The scan fetched 1,148 items, published 611, and marked 307 as recommended.
Published:
Benchmark Radar 第 107 天。你的回测早就知道未来。本轮抓取 1,148 条，发布 611 条，其中 307 条标为推荐。
Published:
Day 106 of Benchmark Radar. The scan fetched 1,111 items, published 496, and marked 283 as recommended.
Published:
Benchmark Radar 第 106 天。真正被测的是运行框架。本轮抓取 1,111 条，发布 496 条，其中 283 条标为推荐。
Published:
Day 105 of Benchmark Radar. The scan fetched 1,322 items, published 599, and marked 205 as recommended.
Published:
Benchmark Radar 第 105 天。配置决定了排名。本轮抓取 1,322 条，发布 599 条，其中 205 条标为推荐。
Published:
Day 104 of Benchmark Radar. The scan fetched 1,462 items, published 744, and marked 307 as recommended.
Published:
Benchmark Radar 第 104 天。评分器会被它评的答案骗过。本轮抓取 1,462 条，发布 744 条，其中 307 条标为推荐。
Published:
Day 103 of Benchmark Radar. The scan fetched 1,627 items, published 831, and marked 444 as recommended.
Published:
Benchmark Radar 第 103 天。基准和软件一样，会积累维护债务。本轮抓取 1,627 条，发布 831 条，其中 444 条标为推荐。
Published:
Day 102 of Benchmark Radar. The scan fetched 1,638 items, published 785, and marked 354 as recommended.
Published:
Benchmark Radar 第 102 天。没有证据的验证器，只是在发表意见。本轮抓取 1,638 条，发布 785 条，其中 354 条标为推荐。
Published:
Day 101 of Benchmark Radar. The scan fetched 1,686 items, published 844, and marked 354 as recommended.
Published:
Benchmark Radar 第 101 天。排行榜排的可能是评测框架。本轮抓取 1,686 条，发布 844 条，其中 354 条标为推荐。
Published:
Day 100 of Benchmark Radar. The scan fetched 1,225 items, published 604, and marked 327 as recommended.
Published:
Benchmark Radar 第 100 天。长不等于深，轮数测的是模型不是任务。本轮抓取 1,225 条，发布 604 条，其中 327 条标为推荐。
Published:
Day 99 of Benchmark Radar. The scan fetched 1,039 items, published 414, and marked 193 as recommended.
Published:
Benchmark Radar 第 99 天。测试时攻击占八成，部署时只占百分之二。本轮抓取 1,039 条，发布 414 条，其中 193 条标为推荐。
Published:
Day 98 of Benchmark Radar. The scan fetched 950 items, published 357, and marked 150 as recommended.
Published:
Benchmark Radar 第 98 天。知道自己在被测的智能体，是另一个智能体。本轮抓取 950 条，发布 357 条，其中 150 条标为推荐。
Published:
Day 97 of Benchmark Radar. The scan fetched 1,202 items, published 517, and marked 238 as recommended.
Published:
Benchmark Radar 第 97 天。每条回复都没问题，整场对话却可以出事。本轮抓取 1,202 条，发布 517 条，其中 238 条标为推荐。
Published:
Day 96 of Benchmark Radar. The scan fetched 1,461 items, published 676, and marked 332 as recommended.
Published:
Benchmark Radar 第 96 天。先检查技能，再让它运行。本轮抓取 1,461 条，发布 676 条，其中 332 条标为推荐。
Published:
Day 95 of Benchmark Radar. The scan fetched 1,308 items, published 568, and marked 250 as recommended.
Published:
Benchmark Radar 第 95 天。会回答空间问题，不等于会在空间里行动。本轮抓取 1,308 条，发布 568 条，其中 250 条标为推荐。
Published:
Day 94 of Benchmark Radar. The scan fetched 1,283 items, published 573, and marked 238 as recommended.
Published:
Benchmark Radar 第 94 天。正确答案，有时就是根本不存在方案。本轮抓取 1,283 条，发布 573 条，其中 238 条标为推荐。
Published:
Day 93 of Benchmark Radar. The scan fetched 1,280 items, published 643, and marked 320 as recommended.
Published:
Benchmark Radar 第 93 天。评判智能体实际改了什么。本轮抓取 1,280 条，发布 643 条，其中 320 条标为推荐。
Published:
Day 92 of Benchmark Radar. The scan fetched 1,048 items, published 464, and marked 211 as recommended.
Published:
Benchmark Radar 第 92 天。经典文档基准的高分，一部分来自错标签和重复数据。本轮抓取 1,048 条，发布 464 条，其中 211 条标为推荐。
Published:
Day 91 of Benchmark Radar. The scan fetched 1,310 items, published 466, and marked 163 as recommended.
Published:
Benchmark Radar 第 91 天。静态图表分数，看不见图表藏起来的东西。本轮抓取 1,310 条，发布 466 条，其中 163 条标为推荐。
Published:
Day 90 of Benchmark Radar. The scan fetched 1,231 items, published 535, and marked 242 as recommended.
Published:
Benchmark Radar 第 90 天。个人记忆需要权限检查，不只是记得准。本轮抓取 1,231 条，发布 535 条，其中 242 条标为推荐。
Published:
Day 89 of Benchmark Radar. The scan fetched 1,332 items, published 630, and marked 289 as recommended.
Published:
Benchmark Radar 第 89 天。工具说真话，也可能漏掉关键证据。本轮抓取 1,332 条，发布 630 条，其中 289 条标为推荐。
Published:
Day 88 of Benchmark Radar. The scan fetched 1,613 items, published 814, and marked 319 as recommended.
Published:
Benchmark Radar 第 88 天。先冻结策略，再测留出期。本轮抓取 1,613 条，发布 814 条，其中 319 条标为推荐。
Published:
Day 87 of Benchmark Radar. The scan fetched 1,692 items, published 884, and marked 391 as recommended.
Published:
Benchmark Radar 第 87 天。任务成功仍需检查授权。本轮抓取 1,692 条，发布 884 条，其中 391 条标为推荐。
Published:
Day 86 of Benchmark Radar. The scan fetched 1,430 items, published 755, and marked 389 as recommended.
Published:
Benchmark Radar 第 86 天。客服智能体还要检查操作过程。本轮抓取 1,430 条，发布 755 条，其中 389 条标为推荐。
Published:
Day 85 of Benchmark Radar. The scan fetched 1,355 items, published 602, and marked 281 as recommended.
Published:
Benchmark Radar 第 85 天。逐条判断证据是否改变。本轮抓取 1,355 条，发布 602 条，其中 281 条标为推荐。
Published:
Day 84 of Benchmark Radar. The scan fetched 1,233 items, published 488, and marked 165 as recommended.
Published:
Benchmark Radar 第 84 天。每个正确答案花多少钱。本轮抓取 1,233 条，发布 488 条，其中 165 条标为推荐。
Published:
Day 83 of Benchmark Radar. The scan fetched 1,562 items, published 749, and marked 246 as recommended.
Published:
Benchmark Radar 第 83 天。安全分数也取决于检测器版本。本轮抓取 1,562 条，发布 749 条，其中 246 条标为推荐。
Published:
Day 82 of Benchmark Radar. The scan fetched 1,672 items, published 846, and marked 346 as recommended.
Published:
Benchmark Radar 第 82 天。输出得分无法证明运行时完整。本轮抓取 1,672 条，发布 846 条，其中 346 条标为推荐。
Published:
Day 81 of Benchmark Radar. The scan fetched 1,711 items, published 899, and marked 425 as recommended.
Published:
Benchmark Radar 第 81 天。去重后仍要按患者划分。本轮抓取 1,711 条，发布 899 条，其中 425 条标为推荐。
Published:
Day 80 of Benchmark Radar. The scan fetched 1,506 items, published 773, and marked 410 as recommended.
Published:
Benchmark Radar 第 80 天。用实验预算评估科研智能体。本轮抓取 1,506 条，发布 773 条，其中 410 条标为推荐。
Published:
Day 79 of Benchmark Radar. The scan fetched 1,476 items, published 624, and marked 323 as recommended.
Published:
Benchmark Radar 第 79 天。标准答案能否从输入推出。本轮抓取 1,476 条，发布 624 条，其中 323 条标为推荐。
Published:
Day 78 of Benchmark Radar. The scan fetched 997 items, published 449, and marked 223 as recommended.
Published:
Benchmark Radar 第 78 天。注入传到了，不等于执行了。本轮抓取 997 条，发布 449 条，其中 223 条标为推荐。
Published:
Day 77 of Benchmark Radar. The scan fetched 1,263 items, published 477, and marked 164 as recommended.
Published:
Benchmark Radar 第 77 天。记忆能否跟上新要求。本轮抓取 1,263 条，发布 477 条，其中 164 条标为推荐。
Published:
Day 76 of Benchmark Radar. The scan fetched 1,316 items, published 547, and marked 234 as recommended.
Published:
Benchmark Radar 第 76 天。测试集要留出完整录制。本轮抓取 1,316 条，发布 547 条，其中 234 条标为推荐。
Published:
Day 75 of Benchmark Radar. The scan fetched 1,338 items, published 616, and marked 281 as recommended.
Published:
Benchmark Radar 第 75 天。分数变了，先查评分器。本轮抓取 1,338 条，发布 616 条，其中 281 条标为推荐。
Published:
Day 74 of Benchmark Radar. The scan fetched 1,326 items, published 604, and marked 274 as recommended.
Published:
Benchmark Radar 第 74 天。安全监控需要完整日志。本轮抓取 1,326 条，发布 604 条，其中 274 条标为推荐。
Published:
Day 73 of Benchmark Radar. The scan fetched 1,657 items, published 833, and marked 376 as recommended.
Published:
Benchmark Radar 第 73 天。不读消息，怎样审计智能体。本轮抓取 1,657 条，发布 833 条，其中 376 条标为推荐。
Published:
Day 72 of Benchmark Radar. The daily snapshot archive has no record for October 6, 2026, so there are no verifiable fetched, published, or recommended counts, or research highlights, for this date.
Published:
Benchmark Radar 第 72 天。每日快照归档中缺少 2026 年 10 月 6 日的记录，因此当天没有可核对的抓取量、发布量、推荐量或研究摘要。
Published:
Day seventy-one of Benchmark Radar. The daily scan published 560 records and recommended 273, led by “Passing the Test You Trained On,” a study of 15 prompt-injection detectors. Scoreboard: 279 stars, 45 forks.
Published:
Benchmark Radar 的第七十一天。每日扫描发布 560 条记录，推荐 273 条，领头的是 《Passing the Test You Trained On》：一项针对 15 个提示注入检测器的研究。记分牌：279 颗星，45 个 fork。
Published:
Day seventy of Benchmark Radar. The daily scan published 484 records and recommended 127, led by the 2026 Cryo-EM Benchmark Dataset, 12 protein structures deposited after every disclosed test set for MICA. Scoreboard: 278 stars, 45 forks.
Published:
Benchmark Radar 的第七十天。每日扫描发布 484 条记录，推荐 127 条，领头的是 2026 Cryo-EM Benchmark Dataset：12 个蛋白结构，沉积时间晚于 MICA 披露的所有测试集。记分牌：278 颗星，45 个 fork。
Published:
Day sixty-nine of Benchmark Radar. The daily scan published 577 records and recommended 174, led by an updated retinal benchmark audit that publishes image-level manifests and per-image predictions. Scoreboard: 275 stars, 45 forks.
Published:
Benchmark Radar 的第六十九天。每日扫描发布 577 条记录，推荐 174 条，领头的是 一项更新后的视网膜基准审计：它公开逐图像的清单和逐图像的预测。记分牌：275 颗星，45 个 fork。
Published:
Day sixty-eight of Benchmark Radar. The daily scan published 668 records and recommended 244, led by Video-Index, a meta-benchmark that audits whether video tests require video understanding. Scoreboard: 272 stars, 45 forks.
Published:
Benchmark Radar 的第六十八天。每日扫描发布 668 条记录，推荐 244 条，领头的是 Video-Index：这个元基准审计视频测试是否真的需要理解视频。记分牌：272 颗星，45 个 fork。
Published:
Day sixty-seven of Benchmark Radar. The daily scan published 688 records and recommended 289, led by “Dating the Model,” a study showing that a hidden current-date line changes measured performance. Scoreboard: 268 stars, 45 forks.
Published:
Benchmark Radar 的第六十七天。每日扫描发布 688 条记录，推荐 289 条，领头的是 《Dating the Model》：这项研究表明，一行隐藏的当前日期会改变测得的表现。记分牌：268 颗星，45 个 fork。
Published:
Day sixty-six of Benchmark Radar. The daily scan published 656 records and recommended 186, led by UserProxyBench, which tests whether the language model playing the user follows its private instructions. Scoreboard: 268 stars, 43 forks.
Published:
Benchmark Radar 的第六十六天。每日扫描发布 656 条记录，推荐 186 条，领头的是 UserProxyBench：它检验扮演用户的语言模型是否遵守它的私有指令。记分牌：268 颗星，43 个 fork。
Published:
Day sixty-five of Benchmark Radar. The daily scan published 758 records and recommended 381, led by TraceDance, which turns real agent deployment traces into benchmarks for undesirable behavior. Scoreboard: 266 stars, 43 forks.
Published:
Benchmark Radar 的第六十五天。每日扫描发布 758 条记录，推荐 381 条，领头的是 TraceDance：它把真实的智能体部署轨迹变成针对不良行为的基准。记分牌：266 颗星，43 个 fork。
Published:
Day sixty-four of Benchmark Radar. The daily scan published 505 records and recommended 233, led by the Finance Agents Benchmark, which packages financial due diligence as an end-to-end agent test. Scoreboard: 264 stars, 42 forks.
Published:
Benchmark Radar 的第六十四天。每日扫描发布 505 条记录，推荐 233 条，领头的是 Finance Agents Benchmark：它把金融尽职调查做成端到端的智能体测试。记分牌：264 颗星，42 个 fork。
Published:
Day sixty-three of Benchmark Radar. The daily scan published 314 records and recommended 106, led by the Kalyvox Voice Benchmark 2026, which scores one voice agent across 240 controlled calls. Scoreboard: 257 stars, 41 forks.
Published:
Benchmark Radar 的第六十三天。每日扫描发布 314 条记录，推荐 106 条，领头的是 Kalyvox Voice Benchmark 2026：它用 240 通受控通话给一个语音智能体打分。记分牌：257 颗星，41 个 fork。
Published:
Day sixty-two of Benchmark Radar. The daily scan published 540 records and recommended 171, led by repo2bench, which turns Python repositories into coding-agent tasks. Scoreboard: 255 stars, 41 forks.
Published:
Benchmark Radar 的第六十二天。每日扫描发布 540 条记录，推荐 171 条，领头的是 repo2bench：它把 Python 仓库转成编程智能体的任务。记分牌：255 颗星，41 个 fork。
Published:
Day sixty-one of Benchmark Radar. The daily scan published 450 records and recommended 198, led by PartHackBench, which tests whether partial-credit scorers reward misleading progress in long-running tool agents. Scoreboard: 252 stars, 41 forks.
Published:
Benchmark Radar 的第六十一天。每日扫描发布 450 条记录，推荐 198 条，领头的是 PartHackBench：它检验部分得分机制是否会奖励长时间运行的工具智能体的误导性进展。记分牌：252 颗星，41 个 fork。
Published:
Day sixty of Benchmark Radar. Benchmark Radar reaches 250 stars on day sixty. The daily scan published 471 records and recommended 199, led by Uncheatable Eval, which scores models on text published after they were trained. Scoreboard: 250 stars, 41 forks.
Published:
Benchmark Radar 的第六十天。第六十天，Benchmark Radar 达到 250 颗星。每日扫描发布 471 条记录，推荐 199 条，领头的是 Uncheatable Eval：它用模型训练之后才发表的文本来打分。记分牌：250 颗星，41 个 fork。
Published:
Day fifty-nine of Benchmark Radar. The daily scan published 510 records and recommended 211, led by a molecular audit that finds hidden duplicates across train and test splits and a benchmark for coding agents whose patches break when merged. Scoreboard: 249 stars, 41 forks.
Published:
Benchmark Radar 的第五十九天。每日扫描发布 510 条记录，推荐 211 条。领头的是一项发现训练集与测试集之间隐藏重复的分子数据审计，以及一个检验编程智能体补丁合并后是否冲突的基准。记分牌：249 颗星，41 个 fork。
Published:
Day fifty-eight of Benchmark Radar. Two fixes clear yesterday’s stop without loosening the check that caused it: Zenodo file batches collapse to one finding, and copied Hugging Face template text no longer passes as a description. Scoreboard: 246 stars, 41 forks.
Published:
Benchmark Radar 的第五十八天。两个修复解除了昨天的停机，却没有放松触发停机的检查：Zenodo 的分文件批量存档合并为一条发现，Hugging Face 模板的复制文本不再冒充描述。记分牌：246 颗星，41 个 fork。
Published:
Day fifty-seven of Benchmark Radar. The daily run failed on purpose. A safety check found 16 records sharing one description and refused to publish them. Scoreboard: 238 stars, 41 forks.
Published:
Benchmark Radar 的第五十七天。每日运行按设计失败了。一项安全检查发现 16 条记录共用同一段描述，拒绝发布。记分牌：238 颗星，41 个 fork。
Published:
Day fifty-six of Benchmark Radar. Contributor winklemad lands two fixes: one makes record-holder credit agree with the record chart, the other keeps every benchmark’s page address unique. Scoreboard: 236 stars, 41 forks.
Published:
Benchmark Radar 的第五十六天。贡献者 winklemad 合并两个修复：一个让纪录保持者的归属与纪录图一致，另一个保证每个基准的页面地址唯一。记分牌：236 颗星，41 个 fork。
Published:
Day fifty-five of Benchmark Radar. The contributor guide gains one rule: start every branch from an up-to-date main. The WeChat group QR code is refreshed, and the daily scan ran on schedule. Scoreboard: 234 stars, 40 forks.
Published:
Benchmark Radar 的第五十五天。贡献者指南新增一条规则：每个分支都从最新的 main 开始。微信群二维码更新，每日扫描按时运行。记分牌：234 颗星，40 个 fork。
Published:
Day fifty-four of Benchmark Radar. The daily scan published 618 records and recommended 224, including a 1.6-million-mask ultrasound release and OverclaimBench, a test of whether coding agents falsely report finished work. Scoreboard: 233 stars, 38 forks.
Published:
Benchmark Radar 的第五十四天。每日扫描发布 618 条记录，推荐 224 条，其中包括一个含 160 万张分割掩码的超声数据发布，以及检验编程智能体是否谎报完工的 OverclaimBench。记分牌：233 颗星，38 个 fork。
Published:
Day fifty-three of Benchmark Radar. The daily scan published 666 records and recommended 261 of them, led by a corrected word-sense benchmark, a stablecoin stress arena, and a ride-hailing agent test. Scoreboard: 229 stars, 36 forks.
Published:
Benchmark Radar 的第五十三天。每日扫描发布 666 条记录，推荐其中 261 条。领头的是一个修正过的词义基准、一个稳定币压力测试场，以及一个网约车智能体测试。记分牌：229 颗星，36 个 fork。
Published:
Day fifty-two of Benchmark Radar. The dataset ships to Hugging Face with automated sync, and seventeen project-only words leave the interface for plain language. Scoreboard: 226 stars, 35 forks.
Published:
Benchmark Radar 的第五十二天。数据集随自动同步上线 Hugging Face，十七处项目黑话退出界面，换成直白语言。记分牌：226 颗星，35 个 fork。
Published:
Day fifty-one of Benchmark Radar. A quiet day with loud groundwork: the daily snapshot recorded while the Hugging Face dataset export goes through validation. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十一天。安静的一天，底下的功夫并不轻：每日快照照常入库，Hugging Face 数据集导出走完验证。记分牌：185 颗星，30 个 fork。
Published:
Day fifty of Benchmark Radar. The paper links point at arXiv everywhere, the README asks readers to upvote on Hugging Face, and a dated laurel badge marks the #1 Paper of the Day on the dashboard. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十天。论文链接全项目指向 arXiv，README 请读者去 Hugging Face 投票，仪表盘挂起带日期的每日第一论文桂冠。记分牌：185 颗星，30 个 fork。
Published:
Day forty-nine of Benchmark Radar. The README starts naming its readers: a researchers-by-institution strip with official logos, built by a script, plus a fresh WeChat group code for daily discussion. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十九天。README 开始点名读者：按机构排列的研究者横幅，官方 logo 由脚本生成；微信群二维码更新，每日基准动态的中文讨论室继续开放。记分牌：185 颗星，30 个 fork。
Published:
Day forty-eight of Benchmark Radar. The dashboard gets a reader-first pass: the homepage banner keeps its words but drops its pixels, benchmark pages answer what readers actually search for, and an About page tells the project’s story on the site. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十八天。仪表盘迎来读者优先改造：首页横幅保留文字、去掉像素占用，基准页围绕读者真实搜索重写，About 页把项目故事讲在站内。记分牌：185 颗星，30 个 fork。
Published:
Day forty-seven of Benchmark Radar. With the arXiv paper live, the citation surfaces get a reader-first refresh: the Cite card collapses APA and BibTeX blocks by default, CITATION.md lands in the repo root, and a Cite link joins the top navigation bar. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十七天。arXiv 论文上线后，引用位置全面转为读者优先：Cite 卡片默认折叠 APA 与 BibTeX，仓库根目录落地 CITATION.md，顶部导航新增 Cite 入口。记分牌：185 颗星，30 个 fork。
Published:
Day forty-six of Benchmark Radar. The technical paper is now on arXiv as arXiv:2609.11115, the v0.11.0 release is published with frozen data and verified hashes, and citation surfaces across the project update to the live arXiv entry. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十六天。技术论文正式登上 arXiv，编号 arXiv:2609.11115；v0.11.0 发布携带冻结数据与经验证哈希；全项目引用位置更新指向在线 arXiv 条目。记分牌：185 颗星，30 个 fork。
Published:
Day forty-five of Benchmark Radar. The paper enters its final preparation phase: content relicensed to CC BY-NC-SA 4.0, CI checks validate frozen paper data, and report tests align with the manuscript. Scoreboard: 180 stars, 32 forks.
Published:
Benchmark Radar 的第四十五天。论文进入最后准备阶段：编辑内容改授 CC BY-NC-SA 4.0，CI 核验冻结论文数据，报告测试与手稿对齐。记分牌：180 颗星，32 个 fork。
Published:
Day forty-four of Benchmark Radar. The paper submodule advances through the citation revision and the source-count revision to the rebuilt PDF. The arXiv submission package assembles. Scoreboard: 175 stars, 31 forks.
Published:
Benchmark Radar 的第四十四天。论文子模块经引文修订、来源计数修订推进到重建 PDF。arXiv 投稿包成型。记分牌：175 颗星，31 个 fork。
Published:
Day forty-three of Benchmark Radar. The discovery cutoff for the technical paper locks at 2026-09-07. Source counts, figure data, and the full-catalog audit pipeline freeze for the v0.11.0 release and arXiv submission. Scoreboard: 170 stars, 30 forks.
Published:
Benchmark Radar 的第四十三天。技术论文的发现截止锁定在 2026-09-07。来源计数、图表数据、全目录审计管道全部冻结，服务 v0.11.0 发布与 arXiv 投稿。记分牌：170 颗星，30 个 fork。
Published:
Day forty-two of Benchmark Radar. The ranking engine for recent releases moves from backend to the dashboard, and the daily radar snapshot captures the first full week of ranked cohorts. Scoreboard: 165 stars, 29 forks.
Published:
Benchmark Radar 的第四十二天。新鲜发布的排行榜引擎从后端走到仪表盘，每日雷达快照记下第一个完整的排序周期。记分牌：165 颗星，29 个 fork。
Published:
Day forty-one of Benchmark Radar. The ranking engine for recent releases shipped its first phase with the data contract written first, and the report audit separated raw best scores from scores measured under the same protocol. Scoreboard: 160 stars, 28 forks.
Published:
Benchmark Radar 的第四十一天。新鲜发布的排行榜引擎第一阶段上线，先把数据契约写在前头；报告审计也把原始最高分和同一协议下测出的分数分开。记分牌：160 颗星，28 个 fork。
Published:
Day forty of Benchmark Radar. The radar started watching the network where most new benchmark papers get their DOI, brought a benchmark back by searching its exact name, and started posting where it was not posting before. Scoreboard: 147 stars, 27 forks.
Published:
Benchmark Radar 的第四十天。雷达开始盯着大多数新基准论文拿到 DOI 的那个网络，靠搜索一个基准的精确名字把它找了回来，并开始向以前没发过的地方发布。记分牌：147 颗星，27 个 fork。
Published:
Day thirty-nine of Benchmark Radar. The daily brief now has a landing page, a full archive, and a feed of its own, and the dashboard navigation got short enough to scan. Scoreboard: 144 stars, 27 forks.
Published:
Benchmark Radar 的第三十九天。每日简报现在有了自己的落地页、完整存档和订阅源，仪表盘导航也短到一眼能扫完。记分牌：144 颗星，27 个 fork。
Published:
Day thirty-eight of Benchmark Radar. The daily brief, the thing the project already wrote every day and never gave a page, now has a blog and a feed. Scoreboard: 137 stars, 26 forks.
Published:
Benchmark Radar 的第三十八天。每日简报，这个项目每天都写、却从没给它一个页面的东西，现在有了博客和订阅源。记分牌：137 颗星，26 个 fork。
Published:
Day thirty-seven of Benchmark Radar. Anyone writing a paper can now copy a citation without leaving the dashboard, the offline setup moved onto the site, and search looks through every collected day by default. Scoreboard: 130 stars, 25 forks.
Published:
Benchmark Radar 的第三十七天。写论文的人现在不用离开仪表盘就能复制引用，离线安装搬到了网站上，搜索默认会翻遍每一个采集过的日子。记分牌：130 颗星，25 个 fork。
Published:
Day thirty-six of Benchmark Radar. The biggest model card yet landed with 37 scores across 34 benchmarks, and search stopped guessing which result answered the question. Scoreboard: 128 stars, 24 forks.
Published:
Benchmark Radar 的第三十六天。迄今最大的一张模型卡落地了，34 个基准上排了 37 个分数；搜索也不再靠猜来回答哪个结果是对的。记分牌：128 颗星，24 个 fork。
Published:
Day thirty-five of Benchmark Radar. A week-old copycat outranked the real project for its own name because search engines could see only four pages here. Today every one of the 1,173 benchmarks the radar tracks has a page of its own. Scoreboard: 123 stars, 23 forks.
Published:
Benchmark Radar 的第三十五天。一个上线才一周的仿站，靠自家名字把真项目压在了搜索结果下面，因为搜索引擎在这里只能看到四个页面。今天，雷达追踪的 1173 个基准，每一个都有了自己独立的页面。记分牌：123 颗星，23 个 fork。
Published:
Day thirty-four of Benchmark Radar. The radar now runs from a command line on your own computer, offline, no dashboard in between, and a coding agent can answer benchmark questions straight from the data. Scoreboard: 121 stars, 23 forks.
Published:
Benchmark Radar 的第三十四天。雷达现在可以从你自己电脑的命令行运行，离线可用，中间不隔仪表盘，编码智能体可以直接从数据里回答基准测试的问题。记分牌：121 颗星，23 个 fork。
Published:
Day thirty-three of Benchmark Radar. The radar learned to find benchmark releases that keyword search would never catch, and a new frontier model card with 14 benchmarks joined the registry. Scoreboard: 115 stars, 21 forks.
Published:
Benchmark Radar 的第三十三天。雷达学会了找到关键词搜索永远抓不到的基准发布，注册表里也进了一张带 14 个基准的新前沿模型卡片。记分牌：115 颗星，21 个 fork。
Published:
Day thirty-two of Benchmark Radar. We fixed data errors that failed silently, taught the ranking to count adoption for releases that ship no binary files, and rewrote the daily briefing so it reads like a person wrote it. Scoreboard: 113 stars, 21 forks.
Published:
Benchmark Radar 的第三十二天。我们修掉了静默失败的数据错误，让排名在发布不附带任何二进制文件时也能统计采纳，还把每日简报改写成像人写的。记分牌：113 颗星，21 个 fork。
Published:
Day thirty-one of Benchmark Radar. Every release card now carries a one-line summary instead of a bare version tag, filing an issue takes at most two required fields, and both READMEs got short enough to read. Scoreboard: 107 stars, 20 forks.
Published:
Benchmark Radar 的第三十一天。每张发布卡片现在都带一行摘要，而不是一个光秃秃的版本号；提交 issue 最多填两个必填项；两份 README 也精简到读得下去。记分牌：107 颗星，20 个 fork。
Published:
Day thirty of Benchmark Radar. Thirty posts in, this log has become a live broadcast of one open-source project trying to reach 1,000 stars, and today the scoreboard reads 86.
Published:
Benchmark Radar 的第三十天。写到今天，这份日志正式变成一场公开直播：一个开源项目如何从零走向 1000 颗星。先报分数：今天记分牌上写着 86。
Published:
Day twenty-nine of Benchmark Radar. We ranked new releases before routine updates, put a 0 to 1 score on a 0 to 100 scale without changing its value, and made the active tab stay active.
Published:
新东西该在第一页，而不是藏在第三页。第二十九天我们做了三件事：让新发布排在例行更新之前，把 0 到 1 的分数按 0 到 100 来显示而不断点，让当前栏目始终看起来是选中的。先说几个词：发布是全新冒出来的基准；更新是对已有条目的改动，例如版本升级或新增分数；0 到 1 与 0 到 100 可以是同一数值的两种刻度。
Published:
Day twenty-eight of Benchmark Radar. We removed a second copy of the data, made the page say one thing to crawlers, and gave the work a citable name.
Published:
别给同一份真相留两个副本，否则它们迟早对不上。第二十八天我们做了三件事：把可生成的数据移出版本库，让页面只对爬虫说一个标题，给作品一个可引用的名字。先说几个词：真相来源是唯一可信的那份源文件，其余都由它生成；h1 是页面的主标题，爬虫期望整页只有一个；i18n 是国际化，让界面按语言显示；引用是你写论文时粘贴的那条参考文献。
Published:
Day twenty-seven of Benchmark Radar. We made the Today list load a page at a time, cut a title in half, and taught 21 benchmarks to show their real names.
Published:
一天 136 条还要一次全画出来，只会让首屏变慢。第二十七天我们做了三件事：让今日列表一页页加载，把标题砍半，让 21 个基准显示真名。先说几个词：分页是一长串内容分多页看完；首包是页面为了快而先加载的小数据包；SEO 是让搜索引擎看懂并收录网站的做法。
Published:
Day twenty-six of Benchmark Radar. We put the ranking where it belongs, made the Back button work, and replaced an invented chart with one we can draw.
Published:
一个以排名命名的页面，却把排名藏在第六屏，这不叫排名页。第二十六天我们做了三件事：让排名站到最前，让返回键能用，把一条编出来的图换成一条能画的线。先说几个词：排名就是按多少排序的榜单；返回键是浏览器左上角让你回到上一页的箭头；前沿线是一条阶梯线，连起截至每个日期为止的最好分数；过期横幅是数据过时时顶部的提示条。
Published:
Day twenty-five of Benchmark Radar. We stopped the radar from ranking itself, put scores on the dates they belong to, and let a benchmark with no score still speak for itself.
Published:
一个会把自己排进榜单的雷达，不配叫雷达。第二十五天我们做了三件事：让排名不再跟自己玩博弈，让每天抓来的分数落在它该在的发布日期上，再让那些还没人评分的基准自己说出身份。先说一个词：排行榜（leaderboard）就是一张公开榜单，谁分数高谁排前面。
Published:
Day twenty-four of Benchmark Radar. We put the results first on the homepage, and we stopped blaming you when a page came up empty.
Published:
今天我们把首页重做了一遍，让它先把结果摆出来，而不是先甩给读者一个空页面。我们还让那些「爬到了但没结果」的来源变得看得见了。先解释两个词：PR 是别人提议改的一段代码，issue 是一个待办或 bug。
Published:
Day twenty-three of Benchmark Radar. We started pulling in other people’s benchmark lists, and we made the leaderboard searchable.
Published:
雷达吸收了外部目录。第二十三天构建了外部目录系统，并在排行榜中新增了基准测试搜索功能。
Published:
Day twenty-two of Benchmark Radar. We locked down data integrity and opened up the community.
Published:
过时运行不再能污染快照数据，微信群正式公开。第二十二天强化了数据完整性并扩展了社区渠道。
Published:
Hi, Koutian here. Day twenty-one was a fix day that touched many parts at once.
Published:
OpenReview 完成了认证接入，宽屏布局得以扩展，手机端的显示问题也修复了。第二十一天是横跨多个模块的修复日。
Published:
Hi, Koutian here. Day twenty taught the radar to speak Chinese and to summarize itself.
Published:
雷达开始说中文了。第二十天交付了完整的 zh 国际化词典、洞察块和完整的社交清单。
Published:
Hi, Koutian here. Day nineteen pushed the radar into maps and satellites.
Published:
雷达扩展到了地理空间和卫星 AI 领域。第十九天为饱和度图添加了领域特定信号和厂商 Logo。
Published:
Hi, Koutian here. Day eighteen was a quiet day, and that was the point.
Published:
你好，我是 Koutian。第十八天，管道自己跑完了，没要任何人插手。这一天，恰恰证明了它稳。
Published:
Hi, Koutian here. Day seventeen was a cleanup day. We fixed what an internal audit found and made the site look finished.
Published:
你好，我是 Koutian。第十七天是加固与美化日。我们没加什么大功能，而是把之前埋的雷排掉，再把界面收拾干净。审计发现的问题都处理完了。
Published:
Some benchmarks were counted wrong, and the dashboards were overwhelming. Day sixteen fixed both.
Published:
你好，我是 Koutian。第十六天，我们修了一个会悄悄骗人的数，还让界面不再一打开就吓到人。
Published:
The radar started posting to social media. Day fifteen built that pipeline and retired the daily GitHub Issue.
Published:
The daily pipeline went from fragile to reliable. Day fourteen added more feeds, hardened the briefing, and turned on Q&A in production.
Published:
The radar now backs up every claim with a link. Day thirteen added evidence grounding, a daily Q&A on the dashboard, and a stack of small fixes.
Published:
The radar now has an RSS feed. Day twelve was a small day, one feature, but it changes how you follow the project.
Published:
KW-Bench got a capability rubric. The community got launch posts. Hi, Koutian here. Day eleven ran on two tracks at once.
Published:
The radar started talking. Hi, Koutian here. Day ten added AI-written briefings, GPT-powered insights, and launch copy for Chinese social platforms.
Published:
雷达开始会自己说话了。第十天我们加了 AI 生成的每日简报、GPT 驱动的洞察，还有面向中文社交平台的发布文案。
Published:
The leaderboard became a research workbench. Hi, Koutian here. Day nine added the adoption frontier, score progression, and saturation charts.
Published:
排行榜以前只是一张排名表。第九天，我们把它改成了能回答问题的研究工作台，加了采纳前沿、分数进展和饱和度可视化。
Published:
A quiet day after the registry explosion. Hi, Koutian here. Day eight settled the pipeline with two precise fixes.
Published:
注册表大爆炸之后，安静的一天。第八天，我们用两次精准修复稳住了流水线。
Published:
Day seven was the biggest single day. Hi, Koutian here. We launched a new leaderboard, grew the registry to the 2026 frontier, and bumped the version to 0.3.0.
Published:
最大的一天。模型卡采纳排行榜上线，注册表扩到了 2026 年前沿，版本号升到 0.3.0。
Published:
The radar now watches Hacker News. Day six added community attention signals and hardened the daily schedule.
Published:
雷达开始听 Hacker News 的动静了。第六天，我们加了社区注意力信号，还加固了每天定时运行的可靠性。
Published:
The trend chart got a real hover card. The landscape report got published. And the agentic benchmark count jumped from 3 to 78.
Published:
趋势图现在有了一个真正的悬浮卡片。全景分析报告发布了。agentic 基准的数量从 3 跳到了 78。
Published:
The radar learned to remember. Day four added historical backfill, staleness detection, and an agentic taxonomy category.
Published:
雷达学会了记忆。第四天，我们加了历史数据回填、过期检测，还有 agentic 这个分类。
Published:
A small day with two precise fixes that made the radar shareable.
Published:
内容不多，但两处精准的修复让雷达变得能分享了。
Published:
AI benchmarks now show up faster than any one researcher can keep up with. So I built a radar that checks for them every day, and shows its work.
Published:
新的 AI 基准冒出来的速度，比任何研究员评估它们的速度都快。所以我建了一个雷达，让「发现新基准」变成一件每天做、过程透明、可以复核的事。先说一个词：基准（benchmark）就是用来考 AI 的一套题，或者一份评测。
Published:
The radar gained memory. Day two added trend maps that span time, plus a fix for benchmarks showing up under many names.
Published:
雷达开始有记忆了。第二天我们建起了累积趋势图，还顺手解决了工件别名的问题。先说一个词：工件（artifact）就是被追踪的每一个具体东西，比如某个基准或数据集。
Published:
Day one of Benchmark Radar. We went from nothing to a working daily dashboard in a single day.
Published:
Benchmark Radar 第一天。我们从零开始，一天之内搭出了一个能跑的累积仪表板。
Published:
Day thirty-nine of Benchmark Radar. The daily brief now has a landing page, a full archive, and a feed of its own, and the dashboard navigation got short enough to scan. Scoreboard: 144 stars, 27 forks.
Published:
Benchmark Radar 的第三十九天。每日简报现在有了自己的落地页、完整存档和订阅源，仪表盘导航也短到一眼能扫完。记分牌：144 颗星，27 个 fork。
Published:
Day thirty-eight of Benchmark Radar. The daily brief, the thing the project already wrote every day and never gave a page, now has a blog and a feed. Scoreboard: 137 stars, 26 forks.
Published:
Benchmark Radar 的第三十八天。每日简报，这个项目每天都写、却从没给它一个页面的东西，现在有了博客和订阅源。记分牌：137 颗星，26 个 fork。
Published:
Day twenty-four of Benchmark Radar. We put the results first on the homepage, and we stopped blaming you when a page came up empty.
Published:
今天我们把首页重做了一遍，让它先把结果摆出来，而不是先甩给读者一个空页面。我们还让那些「爬到了但没结果」的来源变得看得见了。先解释两个词：PR 是别人提议改的一段代码，issue 是一个待办或 bug。
Published:
Day thirty-two of Benchmark Radar. We fixed data errors that failed silently, taught the ranking to count adoption for releases that ship no binary files, and rewrote the daily briefing so it reads like a person wrote it. Scoreboard: 113 stars, 21 forks.
Published:
Benchmark Radar 的第三十二天。我们修掉了静默失败的数据错误，让排名在发布不附带任何二进制文件时也能统计采纳，还把每日简报改写成像人写的。记分牌：113 颗星，21 个 fork。
Published:
Day twenty-nine of Benchmark Radar. We ranked new releases before routine updates, put a 0 to 1 score on a 0 to 100 scale without changing its value, and made the active tab stay active.
Published:
新东西该在第一页，而不是藏在第三页。第二十九天我们做了三件事：让新发布排在例行更新之前，把 0 到 1 的分数按 0 到 100 来显示而不断点，让当前栏目始终看起来是选中的。先说几个词：发布是全新冒出来的基准；更新是对已有条目的改动，例如版本升级或新增分数；0 到 1 与 0 到 100 可以是同一数值的两种刻度。
Published:
The daily pipeline went from fragile to reliable. Day fourteen added more feeds, hardened the briefing, and turned on Q&A in production.
Published:
A junior colleague asked me recently if there is a magical brokerage that does it all, from stock trading to high-yield savings, which reminded me of my own naive expectations when I first started investing.
Published:
这两天有一位好兄弟跑来问我，有没有哪家券商能把炒股、理财和银行存款全包了。
Published:
A street vendor earns a markup. A person who builds a company earns a valuation. Between the two sits one question: who gets other people’s money first.
Published:
Day forty-five of Benchmark Radar. The paper enters its final preparation phase: content relicensed to CC BY-NC-SA 4.0, CI checks validate frozen paper data, and report tests align with the manuscript. Scoreboard: 180 stars, 32 forks.
Published:
Benchmark Radar 的第四十五天。论文进入最后准备阶段：编辑内容改授 CC BY-NC-SA 4.0，CI 核验冻结论文数据，报告测试与手稿对齐。记分牌：180 颗星，32 个 fork。
Published:
The radar started posting to social media. Day fifteen built that pipeline and retired the daily GitHub Issue.
Published:
Day thirty-eight of Benchmark Radar. The daily brief, the thing the project already wrote every day and never gave a page, now has a blog and a feed. Scoreboard: 137 stars, 26 forks.
Published:
Benchmark Radar 的第三十八天。每日简报，这个项目每天都写、却从没给它一个页面的东西，现在有了博客和订阅源。记分牌：137 颗星，26 个 fork。
Published:
Day thirty-seven of Benchmark Radar. Anyone writing a paper can now copy a citation without leaving the dashboard, the offline setup moved onto the site, and search looks through every collected day by default. Scoreboard: 130 stars, 25 forks.
Published:
Benchmark Radar 的第三十七天。写论文的人现在不用离开仪表盘就能复制引用，离线安装搬到了网站上，搜索默认会翻遍每一个采集过的日子。记分牌：130 颗星，25 个 fork。
Published:
Day thirty-four of Benchmark Radar. The radar now runs from a command line on your own computer, offline, no dashboard in between, and a coding agent can answer benchmark questions straight from the data. Scoreboard: 121 stars, 23 forks.
Published:
Benchmark Radar 的第三十四天。雷达现在可以从你自己电脑的命令行运行，离线可用，中间不隔仪表盘，编码智能体可以直接从数据里回答基准测试的问题。记分牌：121 颗星，23 个 fork。
Published:
KW-Bench got a capability rubric. The community got launch posts. Hi, Koutian here. Day eleven ran on two tracks at once.
Published:
Does building a benchmark count as research, or is it “crowdfunding a paper”? Does it create public value or consume public resources? Can it build lasting capabilities, or is it useful only for a startup team trying to show investors its potential?
Published:
Benchmark 到底是在做研究，还是在“众筹一篇论文”？它是在创造公共价值，还是在占用公共资源？它能不能形成长期能力，还是只适合创业团队向投资人展示潜力？
Published:
Day twenty-six of Benchmark Radar. We put the ranking where it belongs, made the Back button work, and replaced an invented chart with one we can draw.
Published:
一个以排名命名的页面，却把排名藏在第六屏，这不叫排名页。第二十六天我们做了三件事：让排名站到最前，让返回键能用，把一条编出来的图换成一条能画的线。先说几个词：排名就是按多少排序的榜单；返回键是浏览器左上角让你回到上一页的箭头；前沿线是一条阶梯线，连起截至每个日期为止的最好分数；过期横幅是数据过时时顶部的提示条。
Published:
Day fifty-nine of Benchmark Radar. The daily scan published 510 records and recommended 211, led by a molecular audit that finds hidden duplicates across train and test splits and a benchmark for coding agents whose patches break when merged. Scoreboard: 249 stars, 41 forks.
Published:
Benchmark Radar 的第五十九天。每日扫描发布 510 条记录，推荐 211 条。领头的是一项发现训练集与测试集之间隐藏重复的分子数据审计，以及一个检验编程智能体补丁合并后是否冲突的基准。记分牌：249 颗星，41 个 fork。
Published:
Hi, Koutian here. Day twenty taught the radar to speak Chinese and to summarize itself.
Published:
雷达开始说中文了。第二十天交付了完整的 zh 国际化词典、洞察块和完整的社交清单。
Published:
Published:
Day forty-seven of Benchmark Radar. With the arXiv paper live, the citation surfaces get a reader-first refresh: the Cite card collapses APA and BibTeX blocks by default, CITATION.md lands in the repo root, and a Cite link joins the top navigation bar. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十七天。arXiv 论文上线后，引用位置全面转为读者优先：Cite 卡片默认折叠 APA 与 BibTeX，仓库根目录落地 CITATION.md，顶部导航新增 Cite 入口。记分牌：185 颗星，30 个 fork。
Published:
Day forty-four of Benchmark Radar. The paper submodule advances through the citation revision and the source-count revision to the rebuilt PDF. The arXiv submission package assembles. Scoreboard: 175 stars, 31 forks.
Published:
Benchmark Radar 的第四十四天。论文子模块经引文修订、来源计数修订推进到重建 PDF。arXiv 投稿包成型。记分牌：175 颗星，31 个 fork。
Published:
Day twenty-eight of Benchmark Radar. We removed a second copy of the data, made the page say one thing to crawlers, and gave the work a citable name.
Published:
别给同一份真相留两个副本，否则它们迟早对不上。第二十八天我们做了三件事：把可生成的数据移出版本库，让页面只对爬虫说一个标题，给作品一个可引用的名字。先说几个词：真相来源是唯一可信的那份源文件，其余都由它生成；h1 是页面的主标题，爬虫期望整页只有一个；i18n 是国际化，让界面按语言显示；引用是你写论文时粘贴的那条参考文献。
Published:
Day thirty-seven of Benchmark Radar. Anyone writing a paper can now copy a citation without leaving the dashboard, the offline setup moved onto the site, and search looks through every collected day by default. Scoreboard: 130 stars, 25 forks.
Published:
Benchmark Radar 的第三十七天。写论文的人现在不用离开仪表盘就能复制引用，离线安装搬到了网站上，搜索默认会翻遍每一个采集过的日子。记分牌：130 颗星，25 个 fork。
Published:
At today’s Silicon Valley alumni “Ask a Question” AI meet-up, what I wanted to talk about wasn’t how cool Claude Code is, but how it actually re-slices real work.
Published:
今天在硅谷校友会的「师问」AI 交流活动上，我想讲的不是 Claude Code 多酷，而是它到底怎么把真实工作重新拆开。
Published:
I ran a health check on my Claude Code setup this week and found 174 custom skills, 124 of which I had never invoked once. They were not failures. Most of them were scaffolding I built for a model that needed it, and then kept long after the model stopped needing it.
Published:
这周，我给自己的 Claude Code 配置做了一次健康检查，发现里面有 174 个自定义 Skill，其中 124 个我一次都没有调用过。它们并不是失败品。大部分只是我为当时需要额外支撑的模型搭起的脚手架，而模型早已不再需要它们，我却一直把它们留到了现在。
Published:
Anthropic’s most ruthless move isn’t polishing every product to perfection before release, but daring to throw out a research preview first, then rapidly getting stronger in front of everyone.
Published:
Anthropic 最狠的地方，不是每次都把产品打磨到完美才发布，而是敢先把一个 research preview 扔出来，然后当着所有人的面快速变强。
Published:
At today’s Silicon Valley alumni “Ask a Question” AI meet-up, what I wanted to talk about wasn’t how cool Claude Code is, but how it actually re-slices real work.
Published:
今天在硅谷校友会的「师问」AI 交流活动上，我想讲的不是 Claude Code 多酷，而是它到底怎么把真实工作重新拆开。
Published:
I was in the middle of setting up a second Cloudflare backup, verifying one dataset and configuring another, when the AI agent helping me flagged something unrelated: while listing what had access to my GitHub account, one authorization stood out as unusually broad. That single flag turned into an hour of reviewing GitHub’s installed-apps list, and it was worth every minute.
Published:
我当时正在配置第二份 Cloudflare 备份——核实一份数据、再搭建另一份——帮我干活的 AI Agent 顺手提醒了我一件不相关的事：在列出哪些东西能访问我的 GitHub 账号时，有一项授权的范围明显大得不正常。就是这一句提醒，让我花了一个多小时去仔细过了一遍 GitHub 的已安装应用列表，而这一个小时非常值得。
Published:
Day fifty-nine of Benchmark Radar. The daily scan published 510 records and recommended 211, led by a molecular audit that finds hidden duplicates across train and test splits and a benchmark for coding agents whose patches break when merged. Scoreboard: 249 stars, 41 forks.
Published:
Benchmark Radar 的第五十九天。每日扫描发布 510 条记录，推荐 211 条。领头的是一项发现训练集与测试集之间隐藏重复的分子数据审计，以及一个检验编程智能体补丁合并后是否冲突的基准。记分牌：249 颗星，41 个 fork。
Published:
Day fifty-four of Benchmark Radar. The daily scan published 618 records and recommended 224, including a 1.6-million-mask ultrasound release and OverclaimBench, a test of whether coding agents falsely report finished work. Scoreboard: 233 stars, 38 forks.
Published:
Benchmark Radar 的第五十四天。每日扫描发布 618 条记录，推荐 224 条，其中包括一个含 160 万张分割掩码的超声数据发布，以及检验编程智能体是否谎报完工的 OverclaimBench。记分牌：233 颗星，38 个 fork。
Published:
Most of biology is averages. You measure a million cells, you get a mean. Eric Greene’s lab figured out how to actually watch one DNA molecule at a time, repair itself, in real time, and that changes what kind of questions biology can ask.
Published:
大部分生物学是平均值。你测一百万个细胞，得到一个均值。Eric Greene 的实验室却想出了怎么一次实时观看一个 DNA 分子自我修复，而这改变了生物学能提出什么样的问题。
Published:
Day forty-nine of Benchmark Radar. The README starts naming its readers: a researchers-by-institution strip with official logos, built by a script, plus a fresh WeChat group code for daily discussion. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十九天。README 开始点名读者：按机构排列的研究者横幅，官方 logo 由脚本生成；微信群二维码更新，每日基准动态的中文讨论室继续开放。记分牌：185 颗星，30 个 fork。
Published:
KW-Bench got a capability rubric. The community got launch posts. Hi, Koutian here. Day eleven ran on two tracks at once.
Published:
Hi, Koutian here. Day seventeen was a cleanup day. We fixed what an internal audit found and made the site look finished.
Published:
你好，我是 Koutian。第十七天是加固与美化日。我们没加什么大功能，而是把之前埋的雷排掉，再把界面收拾干净。审计发现的问题都处理完了。
Published:
I ran a health check on my Claude Code setup this week and found 174 custom skills, 124 of which I had never invoked once. They were not failures. Most of them were scaffolding I built for a model that needed it, and then kept long after the model stopped needing it.
Published:
这周，我给自己的 Claude Code 配置做了一次健康检查，发现里面有 174 个自定义 Skill，其中 124 个我一次都没有调用过。它们并不是失败品。大部分只是我为当时需要额外支撑的模型搭起的脚手架，而模型早已不再需要它们，我却一直把它们留到了现在。
Published:
Day thirty-five of Benchmark Radar. A week-old copycat outranked the real project for its own name because search engines could see only four pages here. Today every one of the 1,173 benchmarks the radar tracks has a page of its own. Scoreboard: 123 stars, 23 forks.
Published:
Benchmark Radar 的第三十五天。一个上线才一周的仿站，靠自家名字把真项目压在了搜索结果下面，因为搜索引擎在这里只能看到四个页面。今天，雷达追踪的 1173 个基准，每一个都有了自己独立的页面。记分牌：123 颗星，23 个 fork。
Published:
Day fifty-six of Benchmark Radar. Contributor winklemad lands two fixes: one makes record-holder credit agree with the record chart, the other keeps every benchmark’s page address unique. Scoreboard: 236 stars, 41 forks.
Published:
Benchmark Radar 的第五十六天。贡献者 winklemad 合并两个修复：一个让纪录保持者的归属与纪录图一致，另一个保证每个基准的页面地址唯一。记分牌：236 颗星，41 个 fork。
Published:
Day fifty-five of Benchmark Radar. The contributor guide gains one rule: start every branch from an up-to-date main. The WeChat group QR code is refreshed, and the daily scan ran on schedule. Scoreboard: 234 stars, 40 forks.
Published:
Benchmark Radar 的第五十五天。贡献者指南新增一条规则：每个分支都从最新的 main 开始。微信群二维码更新，每日扫描按时运行。记分牌：234 颗星，40 个 fork。
Published:
Some benchmarks were counted wrong, and the dashboards were overwhelming. Day sixteen fixed both.
Published:
你好，我是 Koutian。第十六天，我们修了一个会悄悄骗人的数，还让界面不再一打开就吓到人。
Published:
Some books leave behind a few ideas. The Mom Test gave me a method I continue to use: do not ask whether someone likes your idea. Ask what they have already done, how often the problem occurs, what it has cost them, and whether they will commit to a next step.
Published:
有些书读完会留下一些观点，《The Mom Test》留给我的却是一套不断调用的方法：不要问别人喜不喜欢你的想法，要追问他过去真实做过什么、问题多久发生一次、已经付出了什么成本，以及接下来愿不愿意采取行动。
Published:
At Singapore Changi Airport I saw an ad reading “Asia’s Safest Bank,” and my first question wasn’t how impressive it is, it was, who says so?
Published:
我在新加坡樟宜机场看到一块写着「亚洲最安全银行」的广告，脑子里第一个问题不是它有多厉害，而是，谁说的？
Published:
Most of biology is averages. You measure a million cells, you get a mean. Eric Greene’s lab figured out how to actually watch one DNA molecule at a time, repair itself, in real time, and that changes what kind of questions biology can ask.
Published:
大部分生物学是平均值。你测一百万个细胞，得到一个均值。Eric Greene 的实验室却想出了怎么一次实时观看一个 DNA 分子自我修复，而这改变了生物学能提出什么样的问题。
Published:
The radar started talking. Hi, Koutian here. Day ten added AI-written briefings, GPT-powered insights, and launch copy for Chinese social platforms.
Published:
雷达开始会自己说话了。第十天我们加了 AI 生成的每日简报、GPT 驱动的洞察，还有面向中文社交平台的发布文案。
Published:
Day fifty-three of Benchmark Radar. The daily scan published 666 records and recommended 261 of them, led by a corrected word-sense benchmark, a stablecoin stress arena, and a ride-hailing agent test. Scoreboard: 229 stars, 36 forks.
Published:
Benchmark Radar 的第五十三天。每日扫描发布 666 条记录，推荐其中 261 条。领头的是一个修正过的词义基准、一个稳定币压力测试场，以及一个网约车智能体测试。记分牌：229 颗星，36 个 fork。
Published:
Day forty-two of Benchmark Radar. The ranking engine for recent releases moves from backend to the dashboard, and the daily radar snapshot captures the first full week of ranked cohorts. Scoreboard: 165 stars, 29 forks.
Published:
Benchmark Radar 的第四十二天。新鲜发布的排行榜引擎从后端走到仪表盘，每日雷达快照记下第一个完整的排序周期。记分牌：165 颗星，29 个 fork。
Published:
Day thirty-nine of Benchmark Radar. The daily brief now has a landing page, a full archive, and a feed of its own, and the dashboard navigation got short enough to scan. Scoreboard: 144 stars, 27 forks.
Published:
Benchmark Radar 的第三十九天。每日简报现在有了自己的落地页、完整存档和订阅源，仪表盘导航也短到一眼能扫完。记分牌：144 颗星，27 个 fork。
Published:
Day forty-three of Benchmark Radar. The discovery cutoff for the technical paper locks at 2026-09-07. Source counts, figure data, and the full-catalog audit pipeline freeze for the v0.11.0 release and arXiv submission. Scoreboard: 170 stars, 30 forks.
Published:
Benchmark Radar 的第四十三天。技术论文的发现截止锁定在 2026-09-07。来源计数、图表数据、全目录审计管道全部冻结，服务 v0.11.0 发布与 arXiv 投稿。记分牌：170 颗星，30 个 fork。
Published:
Day twenty-two of Benchmark Radar. We locked down data integrity and opened up the community.
Published:
过时运行不再能污染快照数据，微信群正式公开。第二十二天强化了数据完整性并扩展了社区渠道。
Published:
Day thirty-one of Benchmark Radar. Every release card now carries a one-line summary instead of a bare version tag, filing an issue takes at most two required fields, and both READMEs got short enough to read. Scoreboard: 107 stars, 20 forks.
Published:
Benchmark Radar 的第三十一天。每张发布卡片现在都带一行摘要，而不是一个光秃秃的版本号；提交 issue 最多填两个必填项；两份 README 也精简到读得下去。记分牌：107 颗星，20 个 fork。
Published:
Day twenty-eight of Benchmark Radar. We removed a second copy of the data, made the page say one thing to crawlers, and gave the work a citable name.
Published:
别给同一份真相留两个副本，否则它们迟早对不上。第二十八天我们做了三件事：把可生成的数据移出版本库，让页面只对爬虫说一个标题，给作品一个可引用的名字。先说几个词：真相来源是唯一可信的那份源文件，其余都由它生成；h1 是页面的主标题，爬虫期望整页只有一个；i18n 是国际化，让界面按语言显示；引用是你写论文时粘贴的那条参考文献。
Published:
Day fifty-eight of Benchmark Radar. Two fixes clear yesterday’s stop without loosening the check that caused it: Zenodo file batches collapse to one finding, and copied Hugging Face template text no longer passes as a description. Scoreboard: 246 stars, 41 forks.
Published:
Benchmark Radar 的第五十八天。两个修复解除了昨天的停机，却没有放松触发停机的检查：Zenodo 的分文件批量存档合并为一条发现，Hugging Face 模板的复制文本不再冒充描述。记分牌：246 颗星，41 个 fork。
Published:
Day fifty-seven of Benchmark Radar. The daily run failed on purpose. A safety check found 16 records sharing one description and refused to publish them. Scoreboard: 238 stars, 41 forks.
Published:
Benchmark Radar 的第五十七天。每日运行按设计失败了。一项安全检查发现 16 条记录共用同一段描述，拒绝发布。记分牌：238 颗星，41 个 fork。
Published:
Day forty-one of Benchmark Radar. The ranking engine for recent releases shipped its first phase with the data contract written first, and the report audit separated raw best scores from scores measured under the same protocol. Scoreboard: 160 stars, 28 forks.
Published:
Benchmark Radar 的第四十一天。新鲜发布的排行榜引擎第一阶段上线，先把数据契约写在前头；报告审计也把原始最高分和同一协议下测出的分数分开。记分牌：160 颗星，28 个 fork。
Published:
Day thirty-six of Benchmark Radar. The biggest model card yet landed with 37 scores across 34 benchmarks, and search stopped guessing which result answered the question. Scoreboard: 128 stars, 24 forks.
Published:
Benchmark Radar 的第三十六天。迄今最大的一张模型卡落地了，34 个基准上排了 37 个分数；搜索也不再靠猜来回答哪个结果是对的。记分牌：128 颗星，24 个 fork。
Published:
Day thirty-two of Benchmark Radar. We fixed data errors that failed silently, taught the ranking to count adoption for releases that ship no binary files, and rewrote the daily briefing so it reads like a person wrote it. Scoreboard: 113 stars, 21 forks.
Published:
Benchmark Radar 的第三十二天。我们修掉了静默失败的数据错误，让排名在发布不附带任何二进制文件时也能统计采纳，还把每日简报改写成像人写的。记分牌：113 颗星，21 个 fork。
Published:
The radar gained memory. Day two added trend maps that span time, plus a fix for benchmarks showing up under many names.
Published:
雷达开始有记忆了。第二天我们建起了累积趋势图，还顺手解决了工件别名的问题。先说一个词：工件（artifact）就是被追踪的每一个具体东西，比如某个基准或数据集。
Published:
Day fifty-two of Benchmark Radar. The dataset ships to Hugging Face with automated sync, and seventeen project-only words leave the interface for plain language. Scoreboard: 226 stars, 35 forks.
Published:
Benchmark Radar 的第五十二天。数据集随自动同步上线 Hugging Face，十七处项目黑话退出界面，换成直白语言。记分牌：226 颗星，35 个 fork。
Published:
Day fifty-one of Benchmark Radar. A quiet day with loud groundwork: the daily snapshot recorded while the Hugging Face dataset export goes through validation. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十一天。安静的一天，底下的功夫并不轻：每日快照照常入库，Hugging Face 数据集导出走完验证。记分牌：185 颗星，30 个 fork。
Published:
AI benchmarks now show up faster than any one researcher can keep up with. So I built a radar that checks for them every day, and shows its work.
Published:
新的 AI 基准冒出来的速度，比任何研究员评估它们的速度都快。所以我建了一个雷达，让「发现新基准」变成一件每天做、过程透明、可以复核的事。先说一个词：基准（benchmark）就是用来考 AI 的一套题，或者一份评测。
Published:
Day twenty-five of Benchmark Radar. We stopped the radar from ranking itself, put scores on the dates they belong to, and let a benchmark with no score still speak for itself.
Published:
一个会把自己排进榜单的雷达，不配叫雷达。第二十五天我们做了三件事：让排名不再跟自己玩博弈，让每天抓来的分数落在它该在的发布日期上，再让那些还没人评分的基准自己说出身份。先说一个词：排行榜（leaderboard）就是一张公开榜单，谁分数高谁排前面。
Published:
The radar gained memory. Day two added trend maps that span time, plus a fix for benchmarks showing up under many names.
Published:
雷达开始有记忆了。第二天我们建起了累积趋势图，还顺手解决了工件别名的问题。先说一个词：工件（artifact）就是被追踪的每一个具体东西，比如某个基准或数据集。
Published:
| Original source | Tencent Tech; transcript edited | Gu Lingyu; original editors | Xu Qingyang, Su Yang; 2030 annotations and second edit | Koutian Wu |
Published:
原文来源｜腾讯科技；原文整理｜顾翎羽；原文编辑｜徐青阳、苏扬；2030 注释与二次编辑｜Koutian Wu
Published:
I ran a health check on my Claude Code setup this week and found 174 custom skills, 124 of which I had never invoked once. They were not failures. Most of them were scaffolding I built for a model that needed it, and then kept long after the model stopped needing it.
Published:
这周，我给自己的 Claude Code 配置做了一次健康检查，发现里面有 174 个自定义 Skill，其中 124 个我一次都没有调用过。它们并不是失败品。大部分只是我为当时需要额外支撑的模型搭起的脚手架，而模型早已不再需要它们，我却一直把它们留到了现在。
Published:
I spent the last couple of days checking a tempting claim: that Microsoft Flight Simulator is secretly the foundation for a broader Earth digital twin story.
Published:
这两天我在核查一个很容易写得很爽、但也很容易写过头的说法：微软模拟飞行（Microsoft Flight Simulator）是不是某种更大地球数字孪生叙事背后的秘密技术底座。
Published:
Day forty of Benchmark Radar. The radar started watching the network where most new benchmark papers get their DOI, brought a benchmark back by searching its exact name, and started posting where it was not posting before. Scoreboard: 147 stars, 27 forks.
Published:
Benchmark Radar 的第四十天。雷达开始盯着大多数新基准论文拿到 DOI 的那个网络，靠搜索一个基准的精确名字把它找了回来，并开始向以前没发过的地方发布。记分牌：147 颗星，27 个 fork。
Published:
Day thirty-three of Benchmark Radar. The radar learned to find benchmark releases that keyword search would never catch, and a new frontier model card with 14 benchmarks joined the registry. Scoreboard: 115 stars, 21 forks.
Published:
Benchmark Radar 的第三十三天。雷达学会了找到关键词搜索永远抓不到的基准发布，注册表里也进了一张带 14 个基准的新前沿模型卡片。记分牌：115 颗星，21 个 fork。
Published:
Day thirty of Benchmark Radar. Thirty posts in, this log has become a live broadcast of one open-source project trying to reach 1,000 stars, and today the scoreboard reads 86.
Published:
Benchmark Radar 的第三十天。写到今天，这份日志正式变成一场公开直播：一个开源项目如何从零走向 1000 颗星。先报分数：今天记分牌上写着 86。
Published:
Day forty-eight of Benchmark Radar. The dashboard gets a reader-first pass: the homepage banner keeps its words but drops its pixels, benchmark pages answer what readers actually search for, and an About page tells the project’s story on the site. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十八天。仪表盘迎来读者优先改造：首页横幅保留文字、去掉像素占用，基准页围绕读者真实搜索重写，About 页把项目故事讲在站内。记分牌：185 颗星，30 个 fork。
Published:
Day forty-seven of Benchmark Radar. With the arXiv paper live, the citation surfaces get a reader-first refresh: the Cite card collapses APA and BibTeX blocks by default, CITATION.md lands in the repo root, and a Cite link joins the top navigation bar. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十七天。arXiv 论文上线后，引用位置全面转为读者优先：Cite 卡片默认折叠 APA 与 BibTeX，仓库根目录落地 CITATION.md，顶部导航新增 Cite 入口。记分牌：185 颗星，30 个 fork。
Published:
There are only 26 single-letter .ai domains. I went through all of them, one by one, to see who actually builds something and who just squats. The answer: about six run real products, one belongs to Elon Musk, one belongs to Google, one apparently belongs to Apple without a website, and the rest are for-sale pages asking between 1.5 and 500 million dollars. Here is the full alphabet.
Published:
Most VCs and accelerators talk about supporting founders. Kehan Dong specifically supports the kind of founder who started building at 16, when nobody told them they were allowed to, and it turns out that population is dramatically underserved.
Published:
大多数 VC 和孵化器都谈支持创始人。Kehan Dong 专门支持那种 16 岁就开始动手建造、而根本没人告诉过他们可以这么做的创始人，事实证明这个群体被严重忽视了。
Published:
Published:
“The laws of physics look the same in every coordinate system” sounds like a platitude, yet Einstein fought over it for eight years and, halfway through, published a proof that it could not be done.
Published:
Day twenty-four of Benchmark Radar. We put the results first on the homepage, and we stopped blaming you when a page came up empty.
Published:
今天我们把首页重做了一遍，让它先把结果摆出来，而不是先甩给读者一个空页面。我们还让那些「爬到了但没结果」的来源变得看得见了。先解释两个词：PR 是别人提议改的一段代码，issue 是一个待办或 bug。
Published:
A widely shared post describes a brainstorming meeting at a robotics startup in Boston. A slide said “No stupid questions.” A Chinese employee interpreted it as “Do not ask stupid questions,” while American colleagues explained it as “Ask freely; no question will be judged stupid.” A related post claims that “Say it again?” is neutral when someone is not heard, whereas “What did you say?” is hostile.
Published:
一张流传截图讲了这样一件事：一家波士顿机器人创业公司开头脑风暴会，幻灯片写着 “No stupid questions.” 一位中国员工把它理解为“不要问蠢问题”，美国同事却说它的意思是“什么都可以问，没有问题会被当成蠢问题”。另一组讨论又声称：没听清时说 “Say it again?” 很中性，而 “What did you say?” 很不友好。
Published:
The most valuable application of LLMs may not be answering questions, but turning an expert job into a digital work unit that can be executed, reviewed, and audited.
Published:
LLM 最有价值的应用，可能不是回答问题，而是把一项专家工作变成可以执行、复核和审计的数字工作单元。
Published:
Published:
AI benchmarks now show up faster than any one researcher can keep up with. So I built a radar that checks for them every day, and shows its work.
Published:
新的 AI 基准冒出来的速度，比任何研究员评估它们的速度都快。所以我建了一个雷达，让「发现新基准」变成一件每天做、过程透明、可以复核的事。先说一个词：基准（benchmark）就是用来考 AI 的一套题，或者一份评测。
Published:
The most dangerous failure mode of an AI science benchmark is not that it is too hard, it is that it quietly becomes either a coding trick or a guessing game.
Published:
AI 科学 Benchmark 最危险的失败方式，不是它太难，而是它悄悄变成了一道编程技巧题或猜谜题。
Published:
The radar now backs up every claim with a link. Day thirteen added evidence grounding, a daily Q&A on the dashboard, and a stack of small fixes.
Published:
At Singapore Changi Airport I saw an ad reading “Asia’s Safest Bank,” and my first question wasn’t how impressive it is, it was, who says so?
Published:
我在新加坡樟宜机场看到一块写着「亚洲最安全银行」的广告，脑子里第一个问题不是它有多厉害，而是，谁说的？
Published:
Hi, Koutian here. Day seventeen was a cleanup day. We fixed what an internal audit found and made the site look finished.
Published:
你好，我是 Koutian。第十七天是加固与美化日。我们没加什么大功能，而是把之前埋的雷排掉，再把界面收拾干净。审计发现的问题都处理完了。
Published:
Day twenty-three of Benchmark Radar. We started pulling in other people’s benchmark lists, and we made the leaderboard searchable.
Published:
雷达吸收了外部目录。第二十三天构建了外部目录系统，并在排行榜中新增了基准测试搜索功能。
Published:
You apply to Anthropic, land an offer, then ask HR whether they’ll sponsor your PERM, and they say “on hold for now” or change the subject, and you probably stand there frozen.
Published:
你投了 Anthropic，拿到了 offer，然后去问 HR 能不能帮你办 PERM，对方说「目前暂停」或者直接转移话题，你当时大概就愣在原地了。
Published:
Another day of heavy internet surfing, and the biggest lesson was this: there is no investment strategy that lets you lie down and get financially free.
Published:
The radar now backs up every claim with a link. Day thirteen added evidence grounding, a daily Q&A on the dashboard, and a stack of small fixes.
Published:
The daily pipeline went from fragile to reliable. Day fourteen added more feeds, hardened the briefing, and turned on Q&A in production.
Published:
Day twenty-five of Benchmark Radar. We stopped the radar from ranking itself, put scores on the dates they belong to, and let a benchmark with no score still speak for itself.
Published:
一个会把自己排进榜单的雷达，不配叫雷达。第二十五天我们做了三件事：让排名不再跟自己玩博弈，让每天抓来的分数落在它该在的发布日期上，再让那些还没人评分的基准自己说出身份。先说一个词：排行榜（leaderboard）就是一张公开榜单，谁分数高谁排前面。
Published:
A street vendor earns a markup. A person who builds a company earns a valuation. Between the two sits one question: who gets other people’s money first.
Published:
At Singapore Changi Airport I saw an ad reading “Asia’s Safest Bank,” and my first question wasn’t how impressive it is, it was, who says so?
Published:
我在新加坡樟宜机场看到一块写着「亚洲最安全银行」的广告，脑子里第一个问题不是它有多厉害，而是，谁说的？
Published:
A junior colleague asked me recently if there is a magical brokerage that does it all, from stock trading to high-yield savings, which reminded me of my own naive expectations when I first started investing.
Published:
这两天有一位好兄弟跑来问我，有没有哪家券商能把炒股、理财和银行存款全包了。
Published:
My verdict on Air Canada after flying from Singapore to Austin through Vancouver is simple: what a mess.
Published:
The radar learned to remember. Day four added historical backfill, staleness detection, and an agentic taxonomy category.
Published:
雷达学会了记忆。第四天，我们加了历史数据回填、过期检测，还有 agentic 这个分类。
Published:
The most worth-watching thing in 2030 is not how many people are replaced by AI, but which jobs will be reorganized.
Published:
2030 年最值得观察的，不是有多少人被 AI 取代，而是哪些工作会被重新组织。
Published:
The radar started talking. Hi, Koutian here. Day ten added AI-written briefings, GPT-powered insights, and launch copy for Chinese social platforms.
Published:
雷达开始会自己说话了。第十天我们加了 AI 生成的每日简报、GPT 驱动的洞察，还有面向中文社交平台的发布文案。
Published:
A receipt in Singapore showed S$5.50, while a U.S. credit-card account displayed only US$4.27. Another purchase of about S$48 appeared as roughly US$38. A bank-card transit ride also failed to appear immediately as pending.
Published:
在新加坡消费时，一张收据写着 S$5.50，但美国信用卡账户只显示 US$4.27。另一笔约 S$48 的消费，则显示为约 US$38。与此同时，刷银行卡乘坐公共交通后，交易也没有立刻出现在 pending 中。
Published:
“The laws of physics look the same in every coordinate system” sounds like a platitude, yet Einstein fought over it for eight years and, halfway through, published a proof that it could not be done.
Published:
The same way text foundation models ate NLP and image models ate computer vision, someone is going to build the foundation model that eats spatial reasoning. Gengchen Mai is one of the people taking that bet seriously, and his SEAI Lab at UT Austin is one of the places it is being built.
Published:
文本基础模型已经改变了 NLP，图像模型也改变了计算机视觉。空间推理领域同样会出现自己的基础模型。Gengchen Mai 正在研究这一方向，他在 UT Austin 的 SEAI Lab 也开展相关工作。
Published:
Hi, Koutian here. Day nineteen pushed the radar into maps and satellites.
Published:
雷达扩展到了地理空间和卫星 AI 领域。第十九天为饱和度图添加了领域特定信号和厂商 Logo。
Published:
Today, alumnus Pan Jianwei sent his spiritual genes (his name) beyond the Earth.
Published:
今天，校友潘建伟把自己的精神基因（名字）传播到了地球之外。
Published:
I was in the middle of setting up a second Cloudflare backup, verifying one dataset and configuring another, when the AI agent helping me flagged something unrelated: while listing what had access to my GitHub account, one authorization stood out as unusually broad. That single flag turned into an hour of reviewing GitHub’s installed-apps list, and it was worth every minute.
Published:
我当时正在配置第二份 Cloudflare 备份——核实一份数据、再搭建另一份——帮我干活的 AI Agent 顺手提醒了我一件不相关的事：在列出哪些东西能访问我的 GitHub 账号时，有一项授权的范围明显大得不正常。就是这一句提醒，让我花了一个多小时去仔细过了一遍 GitHub 的已安装应用列表，而这一个小时非常值得。
Published:
Going around GitHub opening issues and pull requests is sperm donation. The same mechanism, running as-is.
Published:
在 GitHub 上到处开 issue、提 pull request，就是赛博捐精。同一套机制，原样运行。
Published:
Let me offer a comparison that’s a bit inappropriate but does capture a real structure of open-source collaboration:
Published:
我提一个不太恰当、但确实抓住了开源协作某种结构的比喻：
Published:
Day thirty of Benchmark Radar. Thirty posts in, this log has become a live broadcast of one open-source project trying to reach 1,000 stars, and today the scoreboard reads 86.
Published:
Benchmark Radar 的第三十天。写到今天，这份日志正式变成一场公开直播：一个开源项目如何从零走向 1000 颗星。先报分数：今天记分牌上写着 86。
Published:
The radar now watches Hacker News. Day six added community attention signals and hardened the daily schedule.
Published:
雷达开始听 Hacker News 的动静了。第六天，我们加了社区注意力信号，还加固了每天定时运行的可靠性。
Published:
The radar learned to remember. Day four added historical backfill, staleness detection, and an agentic taxonomy category.
Published:
雷达学会了记忆。第四天，我们加了历史数据回填、过期检测，还有 agentic 这个分类。
Published:
A street vendor earns a markup. A person who builds a company earns a valuation. Between the two sits one question: who gets other people’s money first.
Published:
Day fifty-eight of Benchmark Radar. Two fixes clear yesterday’s stop without loosening the check that caused it: Zenodo file batches collapse to one finding, and copied Hugging Face template text no longer passes as a description. Scoreboard: 246 stars, 41 forks.
Published:
Benchmark Radar 的第五十八天。两个修复解除了昨天的停机，却没有放松触发停机的检查：Zenodo 的分文件批量存档合并为一条发现，Hugging Face 模板的复制文本不再冒充描述。记分牌：246 颗星，41 个 fork。
Published:
Day fifty-two of Benchmark Radar. The dataset ships to Hugging Face with automated sync, and seventeen project-only words leave the interface for plain language. Scoreboard: 226 stars, 35 forks.
Published:
Benchmark Radar 的第五十二天。数据集随自动同步上线 Hugging Face，十七处项目黑话退出界面，换成直白语言。记分牌：226 颗星，35 个 fork。
Published:
Day fifty-one of Benchmark Radar. A quiet day with loud groundwork: the daily snapshot recorded while the Hugging Face dataset export goes through validation. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十一天。安静的一天，底下的功夫并不轻：每日快照照常入库，Hugging Face 数据集导出走完验证。记分牌：185 颗星，30 个 fork。
Published:
Day fifty of Benchmark Radar. The paper links point at arXiv everywhere, the README asks readers to upvote on Hugging Face, and a dated laurel badge marks the #1 Paper of the Day on the dashboard. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十天。论文链接全项目指向 arXiv，README 请读者去 Hugging Face 投票，仪表盘挂起带日期的每日第一论文桂冠。记分牌：185 颗星，30 个 fork。
Published:
Day twenty-five of Benchmark Radar. We stopped the radar from ranking itself, put scores on the dates they belong to, and let a benchmark with no score still speak for itself.
Published:
一个会把自己排进榜单的雷达，不配叫雷达。第二十五天我们做了三件事：让排名不再跟自己玩博弈，让每天抓来的分数落在它该在的发布日期上，再让那些还没人评分的基准自己说出身份。先说一个词：排行榜（leaderboard）就是一张公开榜单，谁分数高谁排前面。
Published:
The story goes like this: I’ve been pondering a question lately—given how insanely advanced technology is today, why do we still have to spend so much time dressing up and managing our personal image?
Published:
故事是这样的，我最近一直在寻思一个问题，既然现在的科技都已经发达到这种地步了，为什么我们这帮人，还是得天天花心思去收拾自己的形象？
Published:
A new paper looked at 739 science Nobel laureates from 1901 to 2023, and the headline number that came out is brutal, at the current rate of progress it would take roughly 600 years before a kid born into a poor country has the same shot at a Nobel as a kid born into a rich one.
Published:
My verdict on Air Canada after flying from Singapore to Austin through Vancouver is simple: what a mess.
Published:
A junior colleague asked me recently if there is a magical brokerage that does it all, from stock trading to high-yield savings, which reminded me of my own naive expectations when I first started investing.
Published:
这两天有一位好兄弟跑来问我，有没有哪家券商能把炒股、理财和银行存款全包了。
Published:
Day twenty-four of Benchmark Radar. We put the results first on the homepage, and we stopped blaming you when a page came up empty.
Published:
今天我们把首页重做了一遍，让它先把结果摆出来，而不是先甩给读者一个空页面。我们还让那些「爬到了但没结果」的来源变得看得见了。先解释两个词：PR 是别人提议改的一段代码，issue 是一个待办或 bug。
Published:
Most academic climate scientists will tell you they care about policy and then publish a paper that no policymaker is ever going to read. Geeta Persad spent four years working at the Union of Concerned Scientists translating climate models for water managers, and she came back to academia knowing exactly what the gap looks like.
Published:
大多数学术气候科学家会告诉你他们在乎政策，然后发表一篇任何决策者都不会读的论文。Geeta Persad 在忧思科学家联盟（Union of Concerned Scientists）花了四年，为水资源管理者翻译气候模型，然后带着对那个差距的清醒认识回到了学术界。
Published:
The clearest advice I recently received from a friend who has visited Japan several times was simple: plan the trip first.
Published:
KW-Bench got a capability rubric. The community got launch posts. Hi, Koutian here. Day eleven ran on two tracks at once.
Published:
The most valuable application of LLMs may not be answering questions, but turning an expert job into a digital work unit that can be executed, reviewed, and audited.
Published:
LLM 最有价值的应用，可能不是回答问题，而是把一项专家工作变成可以执行、复核和审计的数字工作单元。
Published:
Have you ever run into this while using AI?
Published:
你在用 AI 的时候，有没有遇到过这种情况？
Published:
The most valuable application of LLMs may not be answering questions, but turning an expert job into a digital work unit that can be executed, reviewed, and audited.
Published:
LLM 最有价值的应用，可能不是回答问题，而是把一项专家工作变成可以执行、复核和审计的数字工作单元。
Published:
The most worth-watching thing in 2030 is not how many people are replaced by AI, but which jobs will be reorganized.
Published:
2030 年最值得观察的，不是有多少人被 AI 取代，而是哪些工作会被重新组织。
Published:
| Original source | Tencent Tech; transcript edited | Gu Lingyu; original editors | Xu Qingyang, Su Yang; 2030 annotations and second edit | Koutian Wu |
Published:
原文来源｜腾讯科技；原文整理｜顾翎羽；原文编辑｜徐青阳、苏扬；2030 注释与二次编辑｜Koutian Wu
Published:
Prompt Caching Is Not Technical Debt. It’s what the math demands.
Published:
很多人把 prompt caching 看成一个省钱 hack，但我觉得这个判断刚好反了。
Published:
The trend chart got a real hover card. The landscape report got published. And the agentic benchmark count jumped from 3 to 78.
Published:
趋势图现在有了一个真正的悬浮卡片。全景分析报告发布了。agentic 基准的数量从 3 跳到了 78。
Published:
The radar started talking. Hi, Koutian here. Day ten added AI-written briefings, GPT-powered insights, and launch copy for Chinese social platforms.
Published:
雷达开始会自己说话了。第十天我们加了 AI 生成的每日简报、GPT 驱动的洞察，还有面向中文社交平台的发布文案。
Published:
Day one of Benchmark Radar. We went from nothing to a working daily dashboard in a single day.
Published:
Benchmark Radar 第一天。我们从零开始，一天之内搭出了一个能跑的累积仪表板。
Published:
Day twenty-six of Benchmark Radar. We put the ranking where it belongs, made the Back button work, and replaced an invented chart with one we can draw.
Published:
一个以排名命名的页面，却把排名藏在第六屏，这不叫排名页。第二十六天我们做了三件事：让排名站到最前，让返回键能用，把一条编出来的图换成一条能画的线。先说几个词：排名就是按多少排序的榜单；返回键是浏览器左上角让你回到上一页的箭头；前沿线是一条阶梯线，连起截至每个日期为止的最好分数；过期横幅是数据过时时顶部的提示条。
Published:
Day twenty-three of Benchmark Radar. We started pulling in other people’s benchmark lists, and we made the leaderboard searchable.
Published:
雷达吸收了外部目录。第二十三天构建了外部目录系统，并在排行榜中新增了基准测试搜索功能。
Published:
Day seven was the biggest single day. Hi, Koutian here. We launched a new leaderboard, grew the registry to the 2026 frontier, and bumped the version to 0.3.0.
Published:
最大的一天。模型卡采纳排行榜上线，注册表扩到了 2026 年前沿，版本号升到 0.3.0。
Published:
When someone publicly told Steve Jobs that he did not understand technology, Jobs paused, conceded part of the criticism, and asked a more consequential question: what will the customer ultimately gain?
Published:
一个人当众说 Steve Jobs 根本不懂技术，他停了十几秒，没有反击，只问了一件更重要的事，用户最终能得到什么？
Published:
Day forty-five of Benchmark Radar. The paper enters its final preparation phase: content relicensed to CC BY-NC-SA 4.0, CI checks validate frozen paper data, and report tests align with the manuscript. Scoreboard: 180 stars, 32 forks.
Published:
Benchmark Radar 的第四十五天。论文进入最后准备阶段：编辑内容改授 CC BY-NC-SA 4.0，CI 核验冻结论文数据，报告测试与手稿对齐。记分牌：180 颗星，32 个 fork。
Published:
The dark night gave me dark eyes, yet I use them to seek the light.
Published:
黑夜给了我黑色的眼睛，我却用它寻找光明。
Published:
I’ve been watching too many people who got lucky rolling a six six times in a row, standing on stage lecturing us about their dice-rolling strategies.
Published:
这两天看了太多因为连续六次掷出色子六点而功成名就的人，在台上滔滔不绝地分享他们的掷色子心得。
Published:
我曾经以为，自己因为好玩在 LinkedIn 上手动添加几千个人，是假把式，没有任何用。
Published:
The Navier-Stokes fight is not proof that anyone stole anything. It is proof that the person whose work was at stake had no way to check.
Published:
Navier-Stokes 之争并没有证明谁偷了谁的东西。它证明的是：那个把研究押在里面的人，根本无从查证。
Published:
Day fifty-four of Benchmark Radar. The daily scan published 618 records and recommended 224, including a 1.6-million-mask ultrasound release and OverclaimBench, a test of whether coding agents falsely report finished work. Scoreboard: 233 stars, 38 forks.
Published:
Benchmark Radar 的第五十四天。每日扫描发布 618 条记录，推荐 224 条，其中包括一个含 160 万张分割掩码的超声数据发布，以及检验编程智能体是否谎报完工的 OverclaimBench。记分牌：233 颗星，38 个 fork。
Published:
The dark night gave me dark eyes, yet I use them to seek the light.
Published:
黑夜给了我黑色的眼睛，我却用它寻找光明。
Published:
I spent the last couple of days checking a tempting claim: that Microsoft Flight Simulator is secretly the foundation for a broader Earth digital twin story.
Published:
这两天我在核查一个很容易写得很爽、但也很容易写过头的说法：微软模拟飞行（Microsoft Flight Simulator）是不是某种更大地球数字孪生叙事背后的秘密技术底座。
Published:
Microsoft and OpenAI share the same end goal, freeing people, and the only real difference between them is how deep their lever reaches.
Published:
Day sixty of Benchmark Radar. Benchmark Radar reaches 250 stars on day sixty. The daily scan published 471 records and recommended 199, led by Uncheatable Eval, which scores models on text published after they were trained. Scoreboard: 250 stars, 41 forks.
Published:
Benchmark Radar 的第六十天。第六十天，Benchmark Radar 达到 250 颗星。每日扫描发布 471 条记录，推荐 199 条，领头的是 Uncheatable Eval：它用模型训练之后才发表的文本来打分。记分牌：250 颗星，41 个 fork。
Published:
Day thirty-six of Benchmark Radar. The biggest model card yet landed with 37 scores across 34 benchmarks, and search stopped guessing which result answered the question. Scoreboard: 128 stars, 24 forks.
Published:
Benchmark Radar 的第三十六天。迄今最大的一张模型卡落地了，34 个基准上排了 37 个分数；搜索也不再靠猜来回答哪个结果是对的。记分牌：128 颗星，24 个 fork。
Published:
Day thirty-three of Benchmark Radar. The radar learned to find benchmark releases that keyword search would never catch, and a new frontier model card with 14 benchmarks joined the registry. Scoreboard: 115 stars, 21 forks.
Published:
Benchmark Radar 的第三十三天。雷达学会了找到关键词搜索永远抓不到的基准发布，注册表里也进了一张带 14 个基准的新前沿模型卡片。记分牌：115 颗星，21 个 fork。
Published:
Day seven was the biggest single day. Hi, Koutian here. We launched a new leaderboard, grew the registry to the 2026 frontier, and bumped the version to 0.3.0.
Published:
最大的一天。模型卡采纳排行榜上线，注册表扩到了 2026 年前沿，版本号升到 0.3.0。
Published:
One hundred twenty-six thousand eight hundred and eighty-seven dollars is housing-level money in Austin; in UT Austin’s new business-school building, it buys a single parking spot.
Published:
十二万六千八百八十七美元，在 Austin 已经是住房级别的钱；在 UT Austin 的新商学院大楼里，它只对应一个停车位。
Published:
I ended up in the wrong NEXUS line at Vancouver International Airport. By the rules, the mistake was mine. By design, it was an almost predictable wrong turn.
Published:
我走错了温哥华机场的 NEXUS 通道；规则上是我的错，设计上却是一条几乎可以预测的错误路径。
Published:
Day twenty-nine of Benchmark Radar. We ranked new releases before routine updates, put a 0 to 1 score on a 0 to 100 scale without changing its value, and made the active tab stay active.
Published:
新东西该在第一页，而不是藏在第三页。第二十九天我们做了三件事：让新发布排在例行更新之前，把 0 到 1 的分数按 0 到 100 来显示而不断点，让当前栏目始终看起来是选中的。先说几个词：发布是全新冒出来的基准；更新是对已有条目的改动，例如版本升级或新增分数；0 到 1 与 0 到 100 可以是同一数值的两种刻度。
Published:
Day twenty-six of Benchmark Radar. We put the ranking where it belongs, made the Back button work, and replaced an invented chart with one we can draw.
Published:
一个以排名命名的页面，却把排名藏在第六屏，这不叫排名页。第二十六天我们做了三件事：让排名站到最前，让返回键能用，把一条编出来的图换成一条能画的线。先说几个词：排名就是按多少排序的榜单；返回键是浏览器左上角让你回到上一页的箭头；前沿线是一条阶梯线，连起截至每个日期为止的最好分数；过期横幅是数据过时时顶部的提示条。
Published:
When you join a lab, you do not just get a research direction. You inherit a 30-year codebase that is currently running inside the U.S. National Water Model and was on the critical path forecasting Hurricane Harvey. That is what working with Zong-Liang Yang at the Jackson School of Geosciences actually looks like.
Published:
加入实验室以后，你会接过一个研究方向，也会继承一套已有 30 年历史的代码库。它目前运行在美国国家水模型中，也曾处在飓风 Harvey 预报工作的关键路径上。这就是在 Jackson School of Geosciences 与 Zong-Liang Yang 一起工作的真实样子。
Published:
Most land surface models treat a tree like a passive straw. Water comes in at the roots, water leaves at the leaves, end of story. Ashley Matheny’s research basically says no, a tree is an active hydraulic system with storage, capacitance, and a strategy, and if you do not model it that way you are going to be wrong about drought.
Published:
大多数陆地表面模型都把一棵树当成一根被动的吸管。水从根进来，水从叶出去，故事就这么简单。Ashley Matheny 的研究基本上在说：不对，树是一个带有储水、电容和策略的活跃水力系统，如果你不这样建模，你在干旱问题上就会犯错误。
Published:
Most physics undergraduates know Chen Ning Yang as a Nobel laureate and the Y in Yang-Mills. Building him a chronicle made me see something the textbooks gloss over: the most consequential single fact about his life is who his father was, and what that father had set up for him before he was born.
Published:
大多数物理本科生知道杨振宁是诺贝尔奖得主，知道他是 Yang-Mills 里的那个 Yang。而为他建一座年鉴，让我看到了教科书略过的东西：他一生中最有分量的一个事实是谁是他的父亲，以及那个父亲在他出生之前，为他铺好了什么。
Published:
A 1957 Nobel prize cheque, a father who was the first Chinese PhD in number theory, three different monetary regimes, and seventy years of compounding, so how do you put a number on Chen Ning Yang’s family wealth in 2026 without making it up?
Published:
A new paper looked at 739 science Nobel laureates from 1901 to 2023, and the headline number that came out is brutal, at the current rate of progress it would take roughly 600 years before a kid born into a poor country has the same shot at a Nobel as a kid born into a rich one.
Published:
Day twenty-three of Benchmark Radar. We started pulling in other people’s benchmark lists, and we made the leaderboard searchable.
Published:
雷达吸收了外部目录。第二十三天构建了外部目录系统，并在排行榜中新增了基准测试搜索功能。
Published:
A new paper looked at 739 science Nobel laureates from 1901 to 2023, and the headline number that came out is brutal, at the current rate of progress it would take roughly 600 years before a kid born into a poor country has the same shot at a Nobel as a kid born into a rich one.
Published:
一篇新论文把 1901 到 2023 年所有 739 位科学类诺贝尔奖得主从童年扒到现在，得出了一个特别狠的数字，按目前的进步速度，一个出生在低收入国家的孩子要等大约 600 年，才能拿到和富裕国家孩子一样的诺奖机会。
Published:
You can study fluid mechanics in five different countries before you turn 30, work on petroleum reservoirs and tectonophysics and planetary ice on the same week, and somehow end up at a Centennial Chair in Geophysics. Marc Hesse did exactly that, and the through-line is more interesting than any of the individual stops.
Published:
你可以在 30 岁前在五个不同国家学流体力学，同一周里既研究油气储层又研究构造物理和行星冰，最后竟然坐上地球物理学百年讲席。Marc Hesse 就是这么做的，而他背后那条主线比任何一个单独的站点都更有意思。
Published:
Day thirty-four of Benchmark Radar. The radar now runs from a command line on your own computer, offline, no dashboard in between, and a coding agent can answer benchmark questions straight from the data. Scoreboard: 121 stars, 23 forks.
Published:
Benchmark Radar 的第三十四天。雷达现在可以从你自己电脑的命令行运行，离线可用，中间不隔仪表盘，编码智能体可以直接从数据里回答基准测试的问题。记分牌：121 颗星，23 个 fork。
Published:
Day fifty-six of Benchmark Radar. Contributor winklemad lands two fixes: one makes record-holder credit agree with the record chart, the other keeps every benchmark’s page address unique. Scoreboard: 236 stars, 41 forks.
Published:
Benchmark Radar 的第五十六天。贡献者 winklemad 合并两个修复：一个让纪录保持者的归属与纪录图一致，另一个保证每个基准的页面地址唯一。记分牌：236 颗星，41 个 fork。
Published:
Day fifty-five of Benchmark Radar. The contributor guide gains one rule: start every branch from an up-to-date main. The WeChat group QR code is refreshed, and the daily scan ran on schedule. Scoreboard: 234 stars, 40 forks.
Published:
Benchmark Radar 的第五十五天。贡献者指南新增一条规则：每个分支都从最新的 main 开始。微信群二维码更新，每日扫描按时运行。记分牌：234 颗星，40 个 fork。
Published:
Day forty of Benchmark Radar. The radar started watching the network where most new benchmark papers get their DOI, brought a benchmark back by searching its exact name, and started posting where it was not posting before. Scoreboard: 147 stars, 27 forks.
Published:
Benchmark Radar 的第四十天。雷达开始盯着大多数新基准论文拿到 DOI 的那个网络，靠搜索一个基准的精确名字把它找了回来，并开始向以前没发过的地方发布。记分牌：147 颗星，27 个 fork。
Published:
Day thirty-nine of Benchmark Radar. The daily brief now has a landing page, a full archive, and a feed of its own, and the dashboard navigation got short enough to scan. Scoreboard: 144 stars, 27 forks.
Published:
Benchmark Radar 的第三十九天。每日简报现在有了自己的落地页、完整存档和订阅源，仪表盘导航也短到一眼能扫完。记分牌：144 颗星，27 个 fork。
Published:
Day thirty-five of Benchmark Radar. A week-old copycat outranked the real project for its own name because search engines could see only four pages here. Today every one of the 1,173 benchmarks the radar tracks has a page of its own. Scoreboard: 123 stars, 23 forks.
Published:
Benchmark Radar 的第三十五天。一个上线才一周的仿站，靠自家名字把真项目压在了搜索结果下面，因为搜索引擎在这里只能看到四个页面。今天，雷达追踪的 1173 个基准，每一个都有了自己独立的页面。记分牌：123 颗星，23 个 fork。
Published:
Day thirty-four of Benchmark Radar. The radar now runs from a command line on your own computer, offline, no dashboard in between, and a coding agent can answer benchmark questions straight from the data. Scoreboard: 121 stars, 23 forks.
Published:
Benchmark Radar 的第三十四天。雷达现在可以从你自己电脑的命令行运行，离线可用，中间不隔仪表盘，编码智能体可以直接从数据里回答基准测试的问题。记分牌：121 颗星，23 个 fork。
Published:
Day thirty-three of Benchmark Radar. The radar learned to find benchmark releases that keyword search would never catch, and a new frontier model card with 14 benchmarks joined the registry. Scoreboard: 115 stars, 21 forks.
Published:
Benchmark Radar 的第三十三天。雷达学会了找到关键词搜索永远抓不到的基准发布，注册表里也进了一张带 14 个基准的新前沿模型卡片。记分牌：115 颗星，21 个 fork。
Published:
Day thirty-one of Benchmark Radar. Every release card now carries a one-line summary instead of a bare version tag, filing an issue takes at most two required fields, and both READMEs got short enough to read. Scoreboard: 107 stars, 20 forks.
Published:
Benchmark Radar 的第三十一天。每张发布卡片现在都带一行摘要，而不是一个光秃秃的版本号；提交 issue 最多填两个必填项；两份 README 也精简到读得下去。记分牌：107 颗星，20 个 fork。
Published:
Day thirty of Benchmark Radar. Thirty posts in, this log has become a live broadcast of one open-source project trying to reach 1,000 stars, and today the scoreboard reads 86.
Published:
Benchmark Radar 的第三十天。写到今天，这份日志正式变成一场公开直播：一个开源项目如何从零走向 1000 颗星。先报分数：今天记分牌上写着 86。
Published:
AI benchmarks now show up faster than any one researcher can keep up with. So I built a radar that checks for them every day, and shows its work.
Published:
新的 AI 基准冒出来的速度，比任何研究员评估它们的速度都快。所以我建了一个雷达，让「发现新基准」变成一件每天做、过程透明、可以复核的事。先说一个词：基准（benchmark）就是用来考 AI 的一套题，或者一份评测。
Published:
Day one of Benchmark Radar. We went from nothing to a working daily dashboard in a single day.
Published:
Benchmark Radar 第一天。我们从零开始，一天之内搭出了一个能跑的累积仪表板。
Published:
翁家羿在 OpenAI 的职业规划逻辑，是最大化自己名字出现在 OpenAI blog 上的次数。
Published:
Microsoft and OpenAI share the same end goal, freeing people, and the only real difference between them is how deep their lever reaches.
Published:
微软和 OpenAI 在解放人类这个终极目标上高度同构，区别只在杠杆撬到了哪一层。
Published:
Hi, Koutian here. Day twenty-one was a fix day that touched many parts at once.
Published:
OpenReview 完成了认证接入，宽屏布局得以扩展，手机端的显示问题也修复了。第二十一天是横跨多个模块的修复日。
Published:
Hi, Koutian here. Day eighteen was a quiet day, and that was the point.
Published:
你好，我是 Koutian。第十八天，管道自己跑完了，没要任何人插手。这一天，恰恰证明了它稳。
Published:
The trendiest sentence right now is “this tastes too much like AI, we need to de-AI it.”
Published:
现在最时髦的一句话，是“这东西 AI 味太重，得 de-AI”。
Published:
Day forty-four of Benchmark Radar. The paper submodule advances through the citation revision and the source-count revision to the rebuilt PDF. The arXiv submission package assembles. Scoreboard: 175 stars, 31 forks.
Published:
Benchmark Radar 的第四十四天。论文子模块经引文修订、来源计数修订推进到重建 PDF。arXiv 投稿包成型。记分牌：175 颗星，31 个 fork。
Published:
You apply to Anthropic, land an offer, then ask HR whether they’ll sponsor your PERM, and they say “on hold for now” or change the subject, and you probably stand there frozen.
Published:
你投了 Anthropic，拿到了 offer，然后去问 HR 能不能帮你办 PERM，对方说「目前暂停」或者直接转移话题，你当时大概就愣在原地了。
Published:
Day twenty-seven of Benchmark Radar. We made the Today list load a page at a time, cut a title in half, and taught 21 benchmarks to show their real names.
Published:
一天 136 条还要一次全画出来，只会让首屏变慢。第二十七天我们做了三件事：让今日列表一页页加载，把标题砍半，让 21 个基准显示真名。先说几个词：分页是一长串内容分多页看完；首包是页面为了快而先加载的小数据包；SEO 是让搜索引擎看懂并收录网站的做法。
Published:
Day fifty of Benchmark Radar. The paper links point at arXiv everywhere, the README asks readers to upvote on Hugging Face, and a dated laurel badge marks the #1 Paper of the Day on the dashboard. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十天。论文链接全项目指向 arXiv，README 请读者去 Hugging Face 投票，仪表盘挂起带日期的每日第一论文桂冠。记分牌：185 颗星，30 个 fork。
Published:
Day forty-six of Benchmark Radar. The technical paper is now on arXiv as arXiv:2609.11115, the v0.11.0 release is published with frozen data and verified hashes, and citation surfaces across the project update to the live arXiv entry. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十六天。技术论文正式登上 arXiv，编号 arXiv:2609.11115；v0.11.0 发布携带冻结数据与经验证哈希；全项目引用位置更新指向在线 arXiv 条目。记分牌：185 颗星，30 个 fork。
Published:
Day forty-five of Benchmark Radar. The paper enters its final preparation phase: content relicensed to CC BY-NC-SA 4.0, CI checks validate frozen paper data, and report tests align with the manuscript. Scoreboard: 180 stars, 32 forks.
Published:
Benchmark Radar 的第四十五天。论文进入最后准备阶段：编辑内容改授 CC BY-NC-SA 4.0，CI 核验冻结论文数据，报告测试与手稿对齐。记分牌：180 颗星，32 个 fork。
Published:
Day forty-four of Benchmark Radar. The paper submodule advances through the citation revision and the source-count revision to the rebuilt PDF. The arXiv submission package assembles. Scoreboard: 175 stars, 31 forks.
Published:
Benchmark Radar 的第四十四天。论文子模块经引文修订、来源计数修订推进到重建 PDF。arXiv 投稿包成型。记分牌：175 颗星，31 个 fork。
Published:
Day forty-three of Benchmark Radar. The discovery cutoff for the technical paper locks at 2026-09-07. Source counts, figure data, and the full-catalog audit pipeline freeze for the v0.11.0 release and arXiv submission. Scoreboard: 170 stars, 30 forks.
Published:
Benchmark Radar 的第四十三天。技术论文的发现截止锁定在 2026-09-07。来源计数、图表数据、全目录审计管道全部冻结，服务 v0.11.0 发布与 arXiv 投稿。记分牌：170 颗星，30 个 fork。
Published:
Hi, Koutian here. Day twenty-one was a fix day that touched many parts at once.
Published:
OpenReview 完成了认证接入，宽屏布局得以扩展，手机端的显示问题也修复了。第二十一天是横跨多个模块的修复日。
Published:
When you join a lab, you do not just get a research direction. You inherit a 30-year codebase that is currently running inside the U.S. National Water Model and was on the critical path forecasting Hurricane Harvey. That is what working with Zong-Liang Yang at the Jackson School of Geosciences actually looks like.
Published:
加入实验室以后，你会接过一个研究方向，也会继承一套已有 30 年历史的代码库。它目前运行在美国国家水模型中，也曾处在飓风 Harvey 预报工作的关键路径上。这就是在 Jackson School of Geosciences 与 Zong-Liang Yang 一起工作的真实样子。
Published:
You can study fluid mechanics in five different countries before you turn 30, work on petroleum reservoirs and tectonophysics and planetary ice on the same week, and somehow end up at a Centennial Chair in Geophysics. Marc Hesse did exactly that, and the through-line is more interesting than any of the individual stops.
Published:
你可以在 30 岁前在五个不同国家学流体力学，同一周里既研究油气储层又研究构造物理和行星冰，最后竟然坐上地球物理学百年讲席。Marc Hesse 就是这么做的，而他背后那条主线比任何一个单独的站点都更有意思。
Published:
If you ask most people where trees in California get their water in a drought, they will say “the soil.” It turns out a huge fraction of it comes from cracks in the bedrock underneath the soil, and Daniella Rempe is the person who put numbers on it.
Published:
如果你问大多数人，加州在干旱时树从哪里取水，他们会说”土壤”。事实证明，很大一部分水其实来自土壤下方岩石裂缝里的基岩，而 Daniella Rempe 就是把数字放到这件事上的人。
Published:
Most land surface models treat a tree like a passive straw. Water comes in at the roots, water leaves at the leaves, end of story. Ashley Matheny’s research basically says no, a tree is an active hydraulic system with storage, capacitance, and a strategy, and if you do not model it that way you are going to be wrong about drought.
Published:
大多数陆地表面模型都把一棵树当成一根被动的吸管。水从根进来，水从叶出去，故事就这么简单。Ashley Matheny 的研究基本上在说：不对，树是一个带有储水、电容和策略的活跃水力系统，如果你不这样建模，你在干旱问题上就会犯错误。
Published:
来美国读了 PhD 一年半，我得到最好的训练是学会讲好故事。
Published:
In a field where everyone gets pulled toward the next hot thing, Juan Santiago joined Stanford Mechanical Engineering in 1998 and has been there ever since, building one of the most consequential microfluidics labs on the planet. Long-term focus is its own competitive advantage.
Published:
在一个人人都被拽向下一个热门方向的领域里，Juan Santiago 1998 年加入了斯坦福机械工程系，此后一直待在那里，建起了这个星球上最有影响力的微流控实验室之一。长期专注本身就是一种竞争优势。
Published:
“The laws of physics look the same in every coordinate system” sounds like a platitude, yet Einstein fought over it for eight years and, halfway through, published a proof that it could not be done.
Published:
A quiet day after the registry explosion. Hi, Koutian here. Day eight settled the pipeline with two precise fixes.
Published:
注册表大爆炸之后，安静的一天。第八天，我们用两次精准修复稳住了流水线。
Published:
Day fifty-two of Benchmark Radar. The dataset ships to Hugging Face with automated sync, and seventeen project-only words leave the interface for plain language. Scoreboard: 226 stars, 35 forks.
Published:
Benchmark Radar 的第五十二天。数据集随自动同步上线 Hugging Face，十七处项目黑话退出界面，换成直白语言。记分牌：226 颗星，35 个 fork。
Published:
Day forty-nine of Benchmark Radar. The README starts naming its readers: a researchers-by-institution strip with official logos, built by a script, plus a fresh WeChat group code for daily discussion. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十九天。README 开始点名读者：按机构排列的研究者横幅，官方 logo 由脚本生成；微信群二维码更新，每日基准动态的中文讨论室继续开放。记分牌：185 颗星，30 个 fork。
Published:
Day forty-eight of Benchmark Radar. The dashboard gets a reader-first pass: the homepage banner keeps its words but drops its pixels, benchmark pages answer what readers actually search for, and an About page tells the project’s story on the site. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十八天。仪表盘迎来读者优先改造：首页横幅保留文字、去掉像素占用，基准页围绕读者真实搜索重写，About 页把项目故事讲在站内。记分牌：185 颗星，30 个 fork。
Published:
Day forty-seven of Benchmark Radar. With the arXiv paper live, the citation surfaces get a reader-first refresh: the Cite card collapses APA and BibTeX blocks by default, CITATION.md lands in the repo root, and a Cite link joins the top navigation bar. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十七天。arXiv 论文上线后，引用位置全面转为读者优先：Cite 卡片默认折叠 APA 与 BibTeX，仓库根目录落地 CITATION.md，顶部导航新增 Cite 入口。记分牌：185 颗星，30 个 fork。
Published:
Day forty-six of Benchmark Radar. The technical paper is now on arXiv as arXiv:2609.11115, the v0.11.0 release is published with frozen data and verified hashes, and citation surfaces across the project update to the live arXiv entry. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十六天。技术论文正式登上 arXiv，编号 arXiv:2609.11115；v0.11.0 发布携带冻结数据与经验证哈希；全项目引用位置更新指向在线 arXiv 条目。记分牌：185 颗星，30 个 fork。
Published:
Day forty-five of Benchmark Radar. The paper enters its final preparation phase: content relicensed to CC BY-NC-SA 4.0, CI checks validate frozen paper data, and report tests align with the manuscript. Scoreboard: 180 stars, 32 forks.
Published:
Benchmark Radar 的第四十五天。论文进入最后准备阶段：编辑内容改授 CC BY-NC-SA 4.0，CI 核验冻结论文数据，报告测试与手稿对齐。记分牌：180 颗星，32 个 fork。
Published:
Day forty-four of Benchmark Radar. The paper submodule advances through the citation revision and the source-count revision to the rebuilt PDF. The arXiv submission package assembles. Scoreboard: 175 stars, 31 forks.
Published:
Benchmark Radar 的第四十四天。论文子模块经引文修订、来源计数修订推进到重建 PDF。arXiv 投稿包成型。记分牌：175 颗星，31 个 fork。
Published:
Day forty-three of Benchmark Radar. The discovery cutoff for the technical paper locks at 2026-09-07. Source counts, figure data, and the full-catalog audit pipeline freeze for the v0.11.0 release and arXiv submission. Scoreboard: 170 stars, 30 forks.
Published:
Benchmark Radar 的第四十三天。技术论文的发现截止锁定在 2026-09-07。来源计数、图表数据、全目录审计管道全部冻结，服务 v0.11.0 发布与 arXiv 投稿。记分牌：170 颗星，30 个 fork。
Published:
Day forty-two of Benchmark Radar. The ranking engine for recent releases moves from backend to the dashboard, and the daily radar snapshot captures the first full week of ranked cohorts. Scoreboard: 165 stars, 29 forks.
Published:
Benchmark Radar 的第四十二天。新鲜发布的排行榜引擎从后端走到仪表盘，每日雷达快照记下第一个完整的排序周期。记分牌：165 颗星，29 个 fork。
Published:
Day forty-one of Benchmark Radar. The ranking engine for recent releases shipped its first phase with the data contract written first, and the report audit separated raw best scores from scores measured under the same protocol. Scoreboard: 160 stars, 28 forks.
Published:
Benchmark Radar 的第四十一天。新鲜发布的排行榜引擎第一阶段上线，先把数据契约写在前头；报告审计也把原始最高分和同一协议下测出的分数分开。记分牌：160 颗星，28 个 fork。
Published:
Day forty of Benchmark Radar. The radar started watching the network where most new benchmark papers get their DOI, brought a benchmark back by searching its exact name, and started posting where it was not posting before. Scoreboard: 147 stars, 27 forks.
Published:
Benchmark Radar 的第四十天。雷达开始盯着大多数新基准论文拿到 DOI 的那个网络，靠搜索一个基准的精确名字把它找了回来，并开始向以前没发过的地方发布。记分牌：147 颗星，27 个 fork。
Published:
Day thirty-nine of Benchmark Radar. The daily brief now has a landing page, a full archive, and a feed of its own, and the dashboard navigation got short enough to scan. Scoreboard: 144 stars, 27 forks.
Published:
Benchmark Radar 的第三十九天。每日简报现在有了自己的落地页、完整存档和订阅源，仪表盘导航也短到一眼能扫完。记分牌：144 颗星，27 个 fork。
Published:
Day thirty-eight of Benchmark Radar. The daily brief, the thing the project already wrote every day and never gave a page, now has a blog and a feed. Scoreboard: 137 stars, 26 forks.
Published:
Benchmark Radar 的第三十八天。每日简报，这个项目每天都写、却从没给它一个页面的东西，现在有了博客和订阅源。记分牌：137 颗星，26 个 fork。
Published:
Day thirty-seven of Benchmark Radar. Anyone writing a paper can now copy a citation without leaving the dashboard, the offline setup moved onto the site, and search looks through every collected day by default. Scoreboard: 130 stars, 25 forks.
Published:
Benchmark Radar 的第三十七天。写论文的人现在不用离开仪表盘就能复制引用，离线安装搬到了网站上，搜索默认会翻遍每一个采集过的日子。记分牌：130 颗星，25 个 fork。
Published:
Day thirty-six of Benchmark Radar. The biggest model card yet landed with 37 scores across 34 benchmarks, and search stopped guessing which result answered the question. Scoreboard: 128 stars, 24 forks.
Published:
Benchmark Radar 的第三十六天。迄今最大的一张模型卡落地了，34 个基准上排了 37 个分数；搜索也不再靠猜来回答哪个结果是对的。记分牌：128 颗星，24 个 fork。
Published:
Day thirty-five of Benchmark Radar. A week-old copycat outranked the real project for its own name because search engines could see only four pages here. Today every one of the 1,173 benchmarks the radar tracks has a page of its own. Scoreboard: 123 stars, 23 forks.
Published:
Benchmark Radar 的第三十五天。一个上线才一周的仿站，靠自家名字把真项目压在了搜索结果下面，因为搜索引擎在这里只能看到四个页面。今天，雷达追踪的 1173 个基准，每一个都有了自己独立的页面。记分牌：123 颗星，23 个 fork。
Published:
Day thirty-four of Benchmark Radar. The radar now runs from a command line on your own computer, offline, no dashboard in between, and a coding agent can answer benchmark questions straight from the data. Scoreboard: 121 stars, 23 forks.
Published:
Benchmark Radar 的第三十四天。雷达现在可以从你自己电脑的命令行运行，离线可用，中间不隔仪表盘，编码智能体可以直接从数据里回答基准测试的问题。记分牌：121 颗星，23 个 fork。
Published:
Day thirty-three of Benchmark Radar. The radar learned to find benchmark releases that keyword search would never catch, and a new frontier model card with 14 benchmarks joined the registry. Scoreboard: 115 stars, 21 forks.
Published:
Benchmark Radar 的第三十三天。雷达学会了找到关键词搜索永远抓不到的基准发布，注册表里也进了一张带 14 个基准的新前沿模型卡片。记分牌：115 颗星，21 个 fork。
Published:
Day thirty-two of Benchmark Radar. We fixed data errors that failed silently, taught the ranking to count adoption for releases that ship no binary files, and rewrote the daily briefing so it reads like a person wrote it. Scoreboard: 113 stars, 21 forks.
Published:
Benchmark Radar 的第三十二天。我们修掉了静默失败的数据错误，让排名在发布不附带任何二进制文件时也能统计采纳，还把每日简报改写成像人写的。记分牌：113 颗星，21 个 fork。
Published:
Day thirty-one of Benchmark Radar. Every release card now carries a one-line summary instead of a bare version tag, filing an issue takes at most two required fields, and both READMEs got short enough to read. Scoreboard: 107 stars, 20 forks.
Published:
Benchmark Radar 的第三十一天。每张发布卡片现在都带一行摘要，而不是一个光秃秃的版本号；提交 issue 最多填两个必填项；两份 README 也精简到读得下去。记分牌：107 颗星，20 个 fork。
Published:
Day thirty of Benchmark Radar. Thirty posts in, this log has become a live broadcast of one open-source project trying to reach 1,000 stars, and today the scoreboard reads 86.
Published:
Benchmark Radar 的第三十天。写到今天，这份日志正式变成一场公开直播：一个开源项目如何从零走向 1000 颗星。先报分数：今天记分牌上写着 86。
Published:
Day twenty-six of Benchmark Radar. We put the ranking where it belongs, made the Back button work, and replaced an invented chart with one we can draw.
Published:
一个以排名命名的页面，却把排名藏在第六屏，这不叫排名页。第二十六天我们做了三件事：让排名站到最前，让返回键能用，把一条编出来的图换成一条能画的线。先说几个词：排名就是按多少排序的榜单；返回键是浏览器左上角让你回到上一页的箭头；前沿线是一条阶梯线，连起截至每个日期为止的最好分数；过期横幅是数据过时时顶部的提示条。
Published:
Day twenty-five of Benchmark Radar. We stopped the radar from ranking itself, put scores on the dates they belong to, and let a benchmark with no score still speak for itself.
Published:
一个会把自己排进榜单的雷达，不配叫雷达。第二十五天我们做了三件事：让排名不再跟自己玩博弈，让每天抓来的分数落在它该在的发布日期上，再让那些还没人评分的基准自己说出身份。先说一个词：排行榜（leaderboard）就是一张公开榜单，谁分数高谁排前面。
Published:
“The laws of physics look the same in every coordinate system” sounds like a platitude, yet Einstein fought over it for eight years and, halfway through, published a proof that it could not be done.
Published:
The Navier-Stokes fight is not proof that anyone stole anything. It is proof that the person whose work was at stake had no way to check.
Published:
Navier-Stokes 之争并没有证明谁偷了谁的东西。它证明的是：那个把研究押在里面的人，根本无从查证。
Published:
I’ve been watching too many people who got lucky rolling a six six times in a row, standing on stage lecturing us about their dice-rolling strategies.
Published:
这两天看了太多因为连续六次掷出色子六点而功成名就的人，在台上滔滔不绝地分享他们的掷色子心得。
Published:
Some books leave behind a few ideas. The Mom Test gave me a method I continue to use: do not ask whether someone likes your idea. Ask what they have already done, how often the problem occurs, what it has cost them, and whether they will commit to a next step.
Published:
有些书读完会留下一些观点，《The Mom Test》留给我的却是一套不断调用的方法：不要问别人喜不喜欢你的想法，要追问他过去真实做过什么、问题多久发生一次、已经付出了什么成本，以及接下来愿不愿意采取行动。
Published:
When someone publicly told Steve Jobs that he did not understand technology, Jobs paused, conceded part of the criticism, and asked a more consequential question: what will the customer ultimately gain?
Published:
一个人当众说 Steve Jobs 根本不懂技术，他停了十几秒，没有反击，只问了一件更重要的事，用户最终能得到什么？
Published:
The daily pipeline went from fragile to reliable. Day fourteen added more feeds, hardened the briefing, and turned on Q&A in production.
Published:
The most worth-watching thing in 2030 is not how many people are replaced by AI, but which jobs will be reorganized.
Published:
2030 年最值得观察的，不是有多少人被 AI 取代，而是哪些工作会被重新组织。
Published:
Some benchmarks were counted wrong, and the dashboards were overwhelming. Day sixteen fixed both.
Published:
你好，我是 Koutian。第十六天，我们修了一个会悄悄骗人的数，还让界面不再一打开就吓到人。
Published:
The story goes like this: I’ve been pondering a question lately—given how insanely advanced technology is today, why do we still have to spend so much time dressing up and managing our personal image?
Published:
故事是这样的，我最近一直在寻思一个问题，既然现在的科技都已经发达到这种地步了，为什么我们这帮人，还是得天天花心思去收拾自己的形象？
Published:
The radar now backs up every claim with a link. Day thirteen added evidence grounding, a daily Q&A on the dashboard, and a stack of small fixes.
Published:
Day forty-nine of Benchmark Radar. The README starts naming its readers: a researchers-by-institution strip with official logos, built by a script, plus a fresh WeChat group code for daily discussion. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十九天。README 开始点名读者：按机构排列的研究者横幅，官方 logo 由脚本生成；微信群二维码更新，每日基准动态的中文讨论室继续开放。记分牌：185 颗星，30 个 fork。
Published:
Day twenty-two of Benchmark Radar. We locked down data integrity and opened up the community.
Published:
过时运行不再能污染快照数据，微信群正式公开。第二十二天强化了数据完整性并扩展了社区渠道。
Published:
The radar now has an RSS feed. Day twelve was a small day, one feature, but it changes how you follow the project.
Published:
The radar now watches Hacker News. Day six added community attention signals and hardened the daily schedule.
Published:
雷达开始听 Hacker News 的动静了。第六天，我们加了社区注意力信号，还加固了每天定时运行的可靠性。
Published:
Day forty-two of Benchmark Radar. The ranking engine for recent releases moves from backend to the dashboard, and the daily radar snapshot captures the first full week of ranked cohorts. Scoreboard: 165 stars, 29 forks.
Published:
Benchmark Radar 的第四十二天。新鲜发布的排行榜引擎从后端走到仪表盘，每日雷达快照记下第一个完整的排序周期。记分牌：165 颗星，29 个 fork。
Published:
Day forty-one of Benchmark Radar. The ranking engine for recent releases shipped its first phase with the data contract written first, and the report audit separated raw best scores from scores measured under the same protocol. Scoreboard: 160 stars, 28 forks.
Published:
Benchmark Radar 的第四十一天。新鲜发布的排行榜引擎第一阶段上线，先把数据契约写在前头；报告审计也把原始最高分和同一协议下测出的分数分开。记分牌：160 颗星，28 个 fork。
Published:
Day thirty-two of Benchmark Radar. We fixed data errors that failed silently, taught the ranking to count adoption for releases that ship no binary files, and rewrote the daily briefing so it reads like a person wrote it. Scoreboard: 113 stars, 21 forks.
Published:
Benchmark Radar 的第三十二天。我们修掉了静默失败的数据错误，让排名在发布不附带任何二进制文件时也能统计采纳，还把每日简报改写成像人写的。记分牌：113 颗星，21 个 fork。
Published:
Day twenty-nine of Benchmark Radar. We ranked new releases before routine updates, put a 0 to 1 score on a 0 to 100 scale without changing its value, and made the active tab stay active.
Published:
新东西该在第一页，而不是藏在第三页。第二十九天我们做了三件事：让新发布排在例行更新之前，把 0 到 1 的分数按 0 到 100 来显示而不断点，让当前栏目始终看起来是选中的。先说几个词：发布是全新冒出来的基准；更新是对已有条目的改动，例如版本升级或新增分数；0 到 1 与 0 到 100 可以是同一数值的两种刻度。
Published:
What makes a person exceptional? How can we create exceptional things? What separates exceptional work from mediocre work?
Published:
什么样的人是杰出的？我们怎么样才能造出杰出的东西？杰出的东西和平庸的东西有什么区别？
Published:
Does building a benchmark count as research, or is it “crowdfunding a paper”? Does it create public value or consume public resources? Can it build lasting capabilities, or is it useful only for a startup team trying to show investors its potential?
Published:
Benchmark 到底是在做研究，还是在“众筹一篇论文”？它是在创造公共价值，还是在占用公共资源？它能不能形成长期能力，还是只适合创业团队向投资人展示潜力？
Published:
The most dangerous failure mode of an AI science benchmark is not that it is too hard, it is that it quietly becomes either a coding trick or a guessing game.
Published:
AI 科学 Benchmark 最危险的失败方式，不是它太难，而是它悄悄变成了一道编程技巧题或猜谜题。
Published:
Day forty-six of Benchmark Radar. The technical paper is now on arXiv as arXiv:2609.11115, the v0.11.0 release is published with frozen data and verified hashes, and citation surfaces across the project update to the live arXiv entry. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十六天。技术论文正式登上 arXiv，编号 arXiv:2609.11115；v0.11.0 发布携带冻结数据与经验证哈希；全项目引用位置更新指向在线 arXiv 条目。记分牌：185 颗星，30 个 fork。
Published:
The radar now watches Hacker News. Day six added community attention signals and hardened the daily schedule.
Published:
雷达开始听 Hacker News 的动静了。第六天，我们加了社区注意力信号，还加固了每天定时运行的可靠性。
Published:
Researchers who start companies walk in holding a card most founders don’t have.
Published:
研究者去创业，手里其实握着一张很多创业者没有的牌。
Published:
What makes a person exceptional? How can we create exceptional things? What separates exceptional work from mediocre work?
Published:
什么样的人是杰出的？我们怎么样才能造出杰出的东西？杰出的东西和平庸的东西有什么区别？
Published:
A high school student can now submit to Science. The student did not pick the question.
Published:
高中生也能投 Science 了，但题目不是高中生选的。
Published:
The Navier-Stokes fight is not proof that anyone stole anything. It is proof that the person whose work was at stake had no way to check.
Published:
Navier-Stokes 之争并没有证明谁偷了谁的东西。它证明的是：那个把研究押在里面的人，根本无从查证。
Published:
When someone publicly told Steve Jobs that he did not understand technology, Jobs paused, conceded part of the criticism, and asked a more consequential question: what will the customer ultimately gain?
Published:
一个人当众说 Steve Jobs 根本不懂技术，他停了十几秒，没有反击，只问了一件更重要的事，用户最终能得到什么？
Published:
Does building a benchmark count as research, or is it “crowdfunding a paper”? Does it create public value or consume public resources? Can it build lasting capabilities, or is it useful only for a startup team trying to show investors its potential?
Published:
Benchmark 到底是在做研究，还是在“众筹一篇论文”？它是在创造公共价值，还是在占用公共资源？它能不能形成长期能力，还是只适合创业团队向投资人展示潜力？
Published:
A new paper looked at 739 science Nobel laureates from 1901 to 2023, and the headline number that came out is brutal, at the current rate of progress it would take roughly 600 years before a kid born into a poor country has the same shot at a Nobel as a kid born into a rich one.
Published:
Anthropic’s most ruthless move isn’t polishing every product to perfection before release, but daring to throw out a research preview first, then rapidly getting stronger in front of everyone.
Published:
Anthropic 最狠的地方，不是每次都把产品打磨到完美才发布，而是敢先把一个 research preview 扔出来，然后当着所有人的面快速变强。
Published:
A 1957 Nobel prize cheque, a father who was the first Chinese PhD in number theory, three different monetary regimes, and seventy years of compounding, so how do you put a number on Chen Ning Yang’s family wealth in 2026 without making it up?
Published:
The leaderboard became a research workbench. Hi, Koutian here. Day nine added the adoption frontier, score progression, and saturation charts.
Published:
排行榜以前只是一张排名表。第九天，我们把它改成了能回答问题的研究工作台，加了采纳前沿、分数进展和饱和度可视化。
Published:
Hi, Koutian here. Day twenty-one was a fix day that touched many parts at once.
Published:
OpenReview 完成了认证接入，宽屏布局得以扩展，手机端的显示问题也修复了。第二十一天是横跨多个模块的修复日。
Published:
At Singapore Changi Airport I saw an ad reading “Asia’s Safest Bank,” and my first question wasn’t how impressive it is, it was, who says so?
Published:
我在新加坡樟宜机场看到一块写着「亚洲最安全银行」的广告，脑子里第一个问题不是它有多厉害，而是，谁说的？
Published:
Day forty-eight of Benchmark Radar. The dashboard gets a reader-first pass: the homepage banner keeps its words but drops its pixels, benchmark pages answer what readers actually search for, and an About page tells the project’s story on the site. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十八天。仪表盘迎来读者优先改造：首页横幅保留文字、去掉像素占用，基准页围绕读者真实搜索重写，About 页把项目故事讲在站内。记分牌：185 颗星，30 个 fork。
Published:
Day thirty-eight of Benchmark Radar. The daily brief, the thing the project already wrote every day and never gave a page, now has a blog and a feed. Scoreboard: 137 stars, 26 forks.
Published:
Benchmark Radar 的第三十八天。每日简报，这个项目每天都写、却从没给它一个页面的东西，现在有了博客和订阅源。记分牌：137 颗星，26 个 fork。
Published:
Day thirty-five of Benchmark Radar. A week-old copycat outranked the real project for its own name because search engines could see only four pages here. Today every one of the 1,173 benchmarks the radar tracks has a page of its own. Scoreboard: 123 stars, 23 forks.
Published:
Benchmark Radar 的第三十五天。一个上线才一周的仿站，靠自家名字把真项目压在了搜索结果下面，因为搜索引擎在这里只能看到四个页面。今天，雷达追踪的 1173 个基准，每一个都有了自己独立的页面。记分牌：123 颗星，23 个 fork。
Published:
Day twenty-eight of Benchmark Radar. We removed a second copy of the data, made the page say one thing to crawlers, and gave the work a citable name.
Published:
别给同一份真相留两个副本，否则它们迟早对不上。第二十八天我们做了三件事：把可生成的数据移出版本库，让页面只对爬虫说一个标题，给作品一个可引用的名字。先说几个词：真相来源是唯一可信的那份源文件，其余都由它生成；h1 是页面的主标题，爬虫期望整页只有一个；i18n 是国际化，让界面按语言显示；引用是你写论文时粘贴的那条参考文献。
Published:
Day twenty-seven of Benchmark Radar. We made the Today list load a page at a time, cut a title in half, and taught 21 benchmarks to show their real names.
Published:
一天 136 条还要一次全画出来，只会让首屏变慢。第二十七天我们做了三件事：让今日列表一页页加载，把标题砍半，让 21 个基准显示真名。先说几个词：分页是一长串内容分多页看完；首包是页面为了快而先加载的小数据包；SEO 是让搜索引擎看懂并收录网站的做法。
Published:
Hi, Koutian here. Day nineteen pushed the radar into maps and satellites.
Published:
雷达扩展到了地理空间和卫星 AI 领域。第十九天为饱和度图添加了领域特定信号和厂商 Logo。
Published:
Day twenty-nine of Benchmark Radar. We ranked new releases before routine updates, put a 0 to 1 score on a 0 to 100 scale without changing its value, and made the active tab stay active.
Published:
新东西该在第一页，而不是藏在第三页。第二十九天我们做了三件事：让新发布排在例行更新之前，把 0 到 1 的分数按 0 到 100 来显示而不断点，让当前栏目始终看起来是选中的。先说几个词：发布是全新冒出来的基准；更新是对已有条目的改动，例如版本升级或新增分数；0 到 1 与 0 到 100 可以是同一数值的两种刻度。
Published:
Does building a benchmark count as research, or is it “crowdfunding a paper”? Does it create public value or consume public resources? Can it build lasting capabilities, or is it useful only for a startup team trying to show investors its potential?
Published:
Benchmark 到底是在做研究，还是在“众筹一篇论文”？它是在创造公共价值，还是在占用公共资源？它能不能形成长期能力，还是只适合创业团队向投资人展示潜力？
Published:
The most dangerous failure mode of an AI science benchmark is not that it is too hard, it is that it quietly becomes either a coding trick or a guessing game.
Published:
AI 科学 Benchmark 最危险的失败方式，不是它太难，而是它悄悄变成了一道编程技巧题或猜谜题。
Published:
A 1957 Nobel prize cheque, a father who was the first Chinese PhD in number theory, three different monetary regimes, and seventy years of compounding, so how do you put a number on Chen Ning Yang’s family wealth in 2026 without making it up?
Published:
A new paper looked at 739 science Nobel laureates from 1901 to 2023, and the headline number that came out is brutal, at the current rate of progress it would take roughly 600 years before a kid born into a poor country has the same shot at a Nobel as a kid born into a rich one.
Published:
The leaderboard became a research workbench. Hi, Koutian here. Day nine added the adoption frontier, score progression, and saturation charts.
Published:
排行榜以前只是一张排名表。第九天，我们把它改成了能回答问题的研究工作台，加了采纳前沿、分数进展和饱和度可视化。
Published:
Day twenty-five of Benchmark Radar. We stopped the radar from ranking itself, put scores on the dates they belong to, and let a benchmark with no score still speak for itself.
Published:
一个会把自己排进榜单的雷达，不配叫雷达。第二十五天我们做了三件事：让排名不再跟自己玩博弈，让每天抓来的分数落在它该在的发布日期上，再让那些还没人评分的基准自己说出身份。先说一个词：排行榜（leaderboard）就是一张公开榜单，谁分数高谁排前面。
Published:
Day thirty-seven of Benchmark Radar. Anyone writing a paper can now copy a citation without leaving the dashboard, the offline setup moved onto the site, and search looks through every collected day by default. Scoreboard: 130 stars, 25 forks.
Published:
Benchmark Radar 的第三十七天。写论文的人现在不用离开仪表盘就能复制引用，离线安装搬到了网站上，搜索默认会翻遍每一个采集过的日子。记分牌：130 颗星，25 个 fork。
Published:
Day thirty-six of Benchmark Radar. The biggest model card yet landed with 37 scores across 34 benchmarks, and search stopped guessing which result answered the question. Scoreboard: 128 stars, 24 forks.
Published:
Benchmark Radar 的第三十六天。迄今最大的一张模型卡落地了，34 个基准上排了 37 个分数；搜索也不再靠猜来回答哪个结果是对的。记分牌：128 颗星，24 个 fork。
Published:
Day twenty-three of Benchmark Radar. We started pulling in other people’s benchmark lists, and we made the leaderboard searchable.
Published:
雷达吸收了外部目录。第二十三天构建了外部目录系统，并在排行榜中新增了基准测试搜索功能。
Published:
I was in the middle of setting up a second Cloudflare backup, verifying one dataset and configuring another, when the AI agent helping me flagged something unrelated: while listing what had access to my GitHub account, one authorization stood out as unusually broad. That single flag turned into an hour of reviewing GitHub’s installed-apps list, and it was worth every minute.
Published:
我当时正在配置第二份 Cloudflare 备份——核实一份数据、再搭建另一份——帮我干活的 AI Agent 顺手提醒了我一件不相关的事：在列出哪些东西能访问我的 GitHub 账号时，有一项授权的范围明显大得不正常。就是这一句提醒，让我花了一个多小时去仔细过了一遍 GitHub 的已安装应用列表，而这一个小时非常值得。
Published:
A receipt in Singapore showed S$5.50, while a U.S. credit-card account displayed only US$4.27. Another purchase of about S$48 appeared as roughly US$38. A bank-card transit ride also failed to appear immediately as pending.
Published:
在新加坡消费时，一张收据写着 S$5.50，但美国信用卡账户只显示 US$4.27。另一笔约 S$48 的消费，则显示为约 US$38。与此同时，刷银行卡乘坐公共交通后，交易也没有立刻出现在 pending 中。
Published:
At Singapore Changi Airport I saw an ad reading “Asia’s Safest Bank,” and my first question wasn’t how impressive it is, it was, who says so?
Published:
我在新加坡樟宜机场看到一块写着「亚洲最安全银行」的广告，脑子里第一个问题不是它有多厉害，而是，谁说的？
Published:
I found something embarrassing about myself. On many nights I cannot sleep while my AI quota still has room left.
Published:
Day forty of Benchmark Radar. The radar started watching the network where most new benchmark papers get their DOI, brought a benchmark back by searching its exact name, and started posting where it was not posting before. Scoreboard: 147 stars, 27 forks.
Published:
Benchmark Radar 的第四十天。雷达开始盯着大多数新基准论文拿到 DOI 的那个网络，靠搜索一个基准的精确名字把它找了回来，并开始向以前没发过的地方发布。记分牌：147 颗星，27 个 fork。
Published:
Day twenty-two of Benchmark Radar. We locked down data integrity and opened up the community.
Published:
过时运行不再能污染快照数据，微信群正式公开。第二十二天强化了数据完整性并扩展了社区渠道。
Published:
Hi, Koutian here. Day twenty taught the radar to speak Chinese and to summarize itself.
Published:
雷达开始说中文了。第二十天交付了完整的 zh 国际化词典、洞察块和完整的社交清单。
Published:
The radar started posting to social media. Day fifteen built that pipeline and retired the daily GitHub Issue.
Published:
A 1957 Nobel prize cheque, a father who was the first Chinese PhD in number theory, three different monetary regimes, and seventy years of compounding, so how do you put a number on Chen Ning Yang’s family wealth in 2026 without making it up?
Published:
Hi, Koutian here. Day eighteen was a quiet day, and that was the point.
Published:
你好，我是 Koutian。第十八天，管道自己跑完了，没要任何人插手。这一天，恰恰证明了它稳。
Published:
A quiet day after the registry explosion. Hi, Koutian here. Day eight settled the pipeline with two precise fixes.
Published:
注册表大爆炸之后，安静的一天。第八天，我们用两次精准修复稳住了流水线。
Published:
In a field where everyone gets pulled toward the next hot thing, Juan Santiago joined Stanford Mechanical Engineering in 1998 and has been there ever since, building one of the most consequential microfluidics labs on the planet. Long-term focus is its own competitive advantage.
Published:
在一个人人都被拽向下一个热门方向的领域里，Juan Santiago 1998 年加入了斯坦福机械工程系，此后一直待在那里，建起了这个星球上最有影响力的微流控实验室之一。长期专注本身就是一种竞争优势。
Published:
Researchers who start companies walk in holding a card most founders don’t have.
Published:
研究者去创业，手里其实握着一张很多创业者没有的牌。
Published:
A high school student can now submit to Science. The student did not pick the question.
Published:
高中生也能投 Science 了，但题目不是高中生选的。
Published:
Some books leave behind a few ideas. The Mom Test gave me a method I continue to use: do not ask whether someone likes your idea. Ask what they have already done, how often the problem occurs, what it has cost them, and whether they will commit to a next step.
Published:
有些书读完会留下一些观点，《The Mom Test》留给我的却是一套不断调用的方法：不要问别人喜不喜欢你的想法，要追问他过去真实做过什么、问题多久发生一次、已经付出了什么成本，以及接下来愿不愿意采取行动。
Published:
I’ve been watching too many people who got lucky rolling a six six times in a row, standing on stage lecturing us about their dice-rolling strategies.
Published:
这两天看了太多因为连续六次掷出色子六点而功成名就的人，在台上滔滔不绝地分享他们的掷色子心得。
Published:
When someone publicly told Steve Jobs that he did not understand technology, Jobs paused, conceded part of the criticism, and asked a more consequential question: what will the customer ultimately gain?
Published:
一个人当众说 Steve Jobs 根本不懂技术，他停了十几秒，没有反击，只问了一件更重要的事，用户最终能得到什么？
Published:
The radar now has an RSS feed. Day twelve was a small day, one feature, but it changes how you follow the project.
Published:
Some books leave behind a few ideas. The Mom Test gave me a method I continue to use: do not ask whether someone likes your idea. Ask what they have already done, how often the problem occurs, what it has cost them, and whether they will commit to a next step.
Published:
有些书读完会留下一些观点，《The Mom Test》留给我的却是一套不断调用的方法：不要问别人喜不喜欢你的想法，要追问他过去真实做过什么、问题多久发生一次、已经付出了什么成本，以及接下来愿不愿意采取行动。
Published:
I spent the last couple of days checking a tempting claim: that Microsoft Flight Simulator is secretly the foundation for a broader Earth digital twin story.
Published:
这两天我在核查一个很容易写得很爽、但也很容易写过头的说法：微软模拟飞行（Microsoft Flight Simulator）是不是某种更大地球数字孪生叙事背后的秘密技术底座。
Published:
Some books leave behind a few ideas. The Mom Test gave me a method I continue to use: do not ask whether someone likes your idea. Ask what they have already done, how often the problem occurs, what it has cost them, and whether they will commit to a next step.
Published:
有些书读完会留下一些观点，《The Mom Test》留给我的却是一套不断调用的方法：不要问别人喜不喜欢你的想法，要追问他过去真实做过什么、问题多久发生一次、已经付出了什么成本，以及接下来愿不愿意采取行动。
Published:
I’ve been watching too many people who got lucky rolling a six six times in a row, standing on stage lecturing us about their dice-rolling strategies.
Published:
这两天看了太多因为连续六次掷出色子六点而功成名就的人，在台上滔滔不绝地分享他们的掷色子心得。
Published:
The story goes like this: I’ve been pondering a question lately—given how insanely advanced technology is today, why do we still have to spend so much time dressing up and managing our personal image?
Published:
故事是这样的，我最近一直在寻思一个问题，既然现在的科技都已经发达到这种地步了，为什么我们这帮人，还是得天天花心思去收拾自己的形象？
Published:
Day twenty-seven of Benchmark Radar. We made the Today list load a page at a time, cut a title in half, and taught 21 benchmarks to show their real names.
Published:
一天 136 条还要一次全画出来，只会让首屏变慢。第二十七天我们做了三件事：让今日列表一页页加载，把标题砍半，让 21 个基准显示真名。先说几个词：分页是一长串内容分多页看完；首包是页面为了快而先加载的小数据包；SEO 是让搜索引擎看懂并收录网站的做法。
Published:
Day twenty-four of Benchmark Radar. We put the results first on the homepage, and we stopped blaming you when a page came up empty.
Published:
今天我们把首页重做了一遍，让它先把结果摆出来，而不是先甩给读者一个空页面。我们还让那些「爬到了但没结果」的来源变得看得见了。先解释两个词：PR 是别人提议改的一段代码，issue 是一个待办或 bug。
Published:
I ran a health check on my Claude Code setup this week and found 174 custom skills, 124 of which I had never invoked once. They were not failures. Most of them were scaffolding I built for a model that needed it, and then kept long after the model stopped needing it.
Published:
这周，我给自己的 Claude Code 配置做了一次健康检查，发现里面有 174 个自定义 Skill，其中 124 个我一次都没有调用过。它们并不是失败品。大部分只是我为当时需要额外支撑的模型搭起的脚手架，而模型早已不再需要它们，我却一直把它们留到了现在。
Published:
My verdict on Air Canada after flying from Singapore to Austin through Vancouver is simple: what a mess.
Published:
The trend chart got a real hover card. The landscape report got published. And the agentic benchmark count jumped from 3 to 78.
Published:
趋势图现在有了一个真正的悬浮卡片。全景分析报告发布了。agentic 基准的数量从 3 跳到了 78。
Published:
Most physics undergraduates know Chen Ning Yang as a Nobel laureate and the Y in Yang-Mills. Building him a chronicle made me see something the textbooks gloss over: the most consequential single fact about his life is who his father was, and what that father had set up for him before he was born.
Published:
大多数物理本科生知道杨振宁是诺贝尔奖得主，知道他是 Yang-Mills 里的那个 Yang。而为他建一座年鉴，让我看到了教科书略过的东西：他一生中最有分量的一个事实是谁是他的父亲，以及那个父亲在他出生之前，为他铺好了什么。
Published:
Day forty-two of Benchmark Radar. The ranking engine for recent releases moves from backend to the dashboard, and the daily radar snapshot captures the first full week of ranked cohorts. Scoreboard: 165 stars, 29 forks.
Published:
Benchmark Radar 的第四十二天。新鲜发布的排行榜引擎从后端走到仪表盘，每日雷达快照记下第一个完整的排序周期。记分牌：165 颗星，29 个 fork。
Published:
Some benchmarks were counted wrong, and the dashboards were overwhelming. Day sixteen fixed both.
Published:
你好，我是 Koutian。第十六天，我们修了一个会悄悄骗人的数，还让界面不再一打开就吓到人。
Published:
A small day with two precise fixes that made the radar shareable.
Published:
内容不多，但两处精准的修复让雷达变得能分享了。
Published:
Most physics undergraduates know Chen Ning Yang as a Nobel laureate and the Y in Yang-Mills. Building him a chronicle made me see something the textbooks gloss over: the most consequential single fact about his life is who his father was, and what that father had set up for him before he was born.
Published:
大多数物理本科生知道杨振宁是诺贝尔奖得主，知道他是 Yang-Mills 里的那个 Yang。而为他建一座年鉴，让我看到了教科书略过的东西：他一生中最有分量的一个事实是谁是他的父亲，以及那个父亲在他出生之前，为他铺好了什么。
Published:
One hundred twenty-six thousand eight hundred and eighty-seven dollars is housing-level money in Austin; in UT Austin’s new business-school building, it buys a single parking spot.
Published:
十二万六千八百八十七美元，在 Austin 已经是住房级别的钱；在 UT Austin 的新商学院大楼里，它只对应一个停车位。
Published:
When you join a lab, you do not just get a research direction. You inherit a 30-year codebase that is currently running inside the U.S. National Water Model and was on the critical path forecasting Hurricane Harvey. That is what working with Zong-Liang Yang at the Jackson School of Geosciences actually looks like.
Published:
加入实验室以后，你会接过一个研究方向，也会继承一套已有 30 年历史的代码库。它目前运行在美国国家水模型中，也曾处在飓风 Harvey 预报工作的关键路径上。这就是在 Jackson School of Geosciences 与 Zong-Liang Yang 一起工作的真实样子。
Published:
You can study fluid mechanics in five different countries before you turn 30, work on petroleum reservoirs and tectonophysics and planetary ice on the same week, and somehow end up at a Centennial Chair in Geophysics. Marc Hesse did exactly that, and the through-line is more interesting than any of the individual stops.
Published:
你可以在 30 岁前在五个不同国家学流体力学，同一周里既研究油气储层又研究构造物理和行星冰，最后竟然坐上地球物理学百年讲席。Marc Hesse 就是这么做的，而他背后那条主线比任何一个单独的站点都更有意思。
Published:
The same way text foundation models ate NLP and image models ate computer vision, someone is going to build the foundation model that eats spatial reasoning. Gengchen Mai is one of the people taking that bet seriously, and his SEAI Lab at UT Austin is one of the places it is being built.
Published:
文本基础模型已经改变了 NLP，图像模型也改变了计算机视觉。空间推理领域同样会出现自己的基础模型。Gengchen Mai 正在研究这一方向，他在 UT Austin 的 SEAI Lab 也开展相关工作。
Published:
Most academic climate scientists will tell you they care about policy and then publish a paper that no policymaker is ever going to read. Geeta Persad spent four years working at the Union of Concerned Scientists translating climate models for water managers, and she came back to academia knowing exactly what the gap looks like.
Published:
大多数学术气候科学家会告诉你他们在乎政策，然后发表一篇任何决策者都不会读的论文。Geeta Persad 在忧思科学家联盟（Union of Concerned Scientists）花了四年，为水资源管理者翻译气候模型，然后带着对那个差距的清醒认识回到了学术界。
Published:
If you ask most people where trees in California get their water in a drought, they will say “the soil.” It turns out a huge fraction of it comes from cracks in the bedrock underneath the soil, and Daniella Rempe is the person who put numbers on it.
Published:
如果你问大多数人，加州在干旱时树从哪里取水，他们会说”土壤”。事实证明，很大一部分水其实来自土壤下方岩石裂缝里的基岩，而 Daniella Rempe 就是把数字放到这件事上的人。
Published:
Most land surface models treat a tree like a passive straw. Water comes in at the roots, water leaves at the leaves, end of story. Ashley Matheny’s research basically says no, a tree is an active hydraulic system with storage, capacitance, and a strategy, and if you do not model it that way you are going to be wrong about drought.
Published:
大多数陆地表面模型都把一棵树当成一根被动的吸管。水从根进来，水从叶出去，故事就这么简单。Ashley Matheny 的研究基本上在说：不对，树是一个带有储水、电容和策略的活跃水力系统，如果你不这样建模，你在干旱问题上就会犯错误。
Published:
这次 Texas ML Symposium 上最让我记住的，是一场讲怎么扩展语言模型的报告，不是怎么把模型做大。
Published:
Day forty-eight of Benchmark Radar. The dashboard gets a reader-first pass: the homepage banner keeps its words but drops its pixels, benchmark pages answer what readers actually search for, and an About page tells the project’s story on the site. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十八天。仪表盘迎来读者优先改造：首页横幅保留文字、去掉像素占用，基准页围绕读者真实搜索重写，About 页把项目故事讲在站内。记分牌：185 颗星，30 个 fork。
Published:
Day forty-seven of Benchmark Radar. With the arXiv paper live, the citation surfaces get a reader-first refresh: the Cite card collapses APA and BibTeX blocks by default, CITATION.md lands in the repo root, and a Cite link joins the top navigation bar. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十七天。arXiv 论文上线后，引用位置全面转为读者优先：Cite 卡片默认折叠 APA 与 BibTeX，仓库根目录落地 CITATION.md，顶部导航新增 Cite 入口。记分牌：185 颗星，30 个 fork。
Published:
Hi, Koutian here. Day twenty taught the radar to speak Chinese and to summarize itself.
Published:
雷达开始说中文了。第二十天交付了完整的 zh 国际化词典、洞察块和完整的社交清单。
Published:
Hi, Koutian here. Day seventeen was a cleanup day. We fixed what an internal audit found and made the site look finished.
Published:
你好，我是 Koutian。第十七天是加固与美化日。我们没加什么大功能，而是把之前埋的雷排掉，再把界面收拾干净。审计发现的问题都处理完了。
Published:
A small day with two precise fixes that made the radar shareable.
Published:
内容不多，但两处精准的修复让雷达变得能分享了。
Published:
A street vendor earns a markup. A person who builds a company earns a valuation. Between the two sits one question: who gets other people’s money first.
Published:
A street vendor earns a markup. A person who builds a company earns a valuation. Between the two sits one question: who gets other people’s money first.
Published:
Day seven was the biggest single day. Hi, Koutian here. We launched a new leaderboard, grew the registry to the 2026 frontier, and bumped the version to 0.3.0.
Published:
最大的一天。模型卡采纳排行榜上线，注册表扩到了 2026 年前沿，版本号升到 0.3.0。
Published:
Hi, Koutian here. Day nineteen pushed the radar into maps and satellites.
Published:
雷达扩展到了地理空间和卫星 AI 领域。第十九天为饱和度图添加了领域特定信号和厂商 Logo。
Published:
The leaderboard became a research workbench. Hi, Koutian here. Day nine added the adoption frontier, score progression, and saturation charts.
Published:
排行榜以前只是一张排名表。第九天，我们把它改成了能回答问题的研究工作台，加了采纳前沿、分数进展和饱和度可视化。
Published:
Published:
Day twenty-two of Benchmark Radar. We locked down data integrity and opened up the community.
Published:
过时运行不再能污染快照数据，微信群正式公开。第二十二天强化了数据完整性并扩展了社区渠道。
Published:
The radar started posting to social media. Day fifteen built that pipeline and retired the daily GitHub Issue.
Published:
A 1957 Nobel prize cheque, a father who was the first Chinese PhD in number theory, three different monetary regimes, and seventy years of compounding, so how do you put a number on Chen Ning Yang’s family wealth in 2026 without making it up?
Published:
A new paper looked at 739 science Nobel laureates from 1901 to 2023, and the headline number that came out is brutal, at the current rate of progress it would take roughly 600 years before a kid born into a poor country has the same shot at a Nobel as a kid born into a rich one.
Published:
The trendiest sentence right now is “this tastes too much like AI, we need to de-AI it.”
Published:
现在最时髦的一句话，是“这东西 AI 味太重，得 de-AI”。
Published:
Most enterprise security companies show up, ride one trend, and disappear in the next infrastructure cycle. Sean Xiang has been building Bloombase since January 2012, and the company has somehow been on the right side of every major infrastructure shift since, including the current AI accelerator era. That is not luck. That is a thesis.
Published:
大多数企业安全公司冒个泡、赶一波趋势，然后在下一个基础设施周期里消失。Sean Xiang 从 2012 年 1 月起就在打造 Bloombase，这家公司却阴差阳错站到了此后每一波重大基础设施转变的正确一边，包括当下的 AI 加速器时代。那不是运气，那是一套论点。
Published:
Most VCs and accelerators talk about supporting founders. Kehan Dong specifically supports the kind of founder who started building at 16, when nobody told them they were allowed to, and it turns out that population is dramatically underserved.
Published:
大多数 VC 和孵化器都谈支持创始人。Kehan Dong 专门支持那种 16 岁就开始动手建造、而根本没人告诉过他们可以这么做的创始人，事实证明这个群体被严重忽视了。
Published:
Rereading my old journals today, from nuclear fusion in physics to the games humans play, I keep finding the same thing. Whether it is a startup or a whole industry, what matters is a force that rises on its own and then keeps itself going.
Published:
I ended up in the wrong NEXUS line at Vancouver International Airport. By the rules, the mistake was mine. By design, it was an almost predictable wrong turn.
Published:
我走错了温哥华机场的 NEXUS 通道；规则上是我的错，设计上却是一条几乎可以预测的错误路径。
Published:
Beyond the wayfinding problem that led me into the wrong NEXUS line, my connection at Vancouver International Airport left me with several observations unrelated to signs. I noticed an older-looking mix of people, a departures area that offered almost nowhere to work, and a visible contrast between people in Vancouver and Texas. This post records those observations. The full account of the NEXUS incident is in How I “Reasonably” Ended Up in the Wrong NEXUS Line at YVR.
Published:
除了那次走错 NEXUS 通道的导视问题，这次在温哥华机场（YVR）转机还让我观察到几件和路标无关、但同样值得记录的事情：明显偏大的现场年龄结构、几乎不为办公设计的候机空间，以及温哥华与得州人群景观的直观差异。这篇笔记单独记录这些观察；如果你想看我是怎么在 NEXUS 通道走错路的，可以读我在 YVR”合理地”走错了 NEXUS。
Published:
My verdict on Air Canada after flying from Singapore to Austin through Vancouver is simple: what a mess.
Published:
这次从新加坡经温哥华飞回奥斯汀，我对 Air Canada 的评价很简单：太拉了。
Published:
A 1957 Nobel prize cheque, a father who was the first Chinese PhD in number theory, three different monetary regimes, and seventy years of compounding, so how do you put a number on Chen Ning Yang’s family wealth in 2026 without making it up?
Published:
Most physics undergraduates know Chen Ning Yang as a Nobel laureate and the Y in Yang-Mills. Building him a chronicle made me see something the textbooks gloss over: the most consequential single fact about his life is who his father was, and what that father had set up for him before he was born.
Published:
大多数物理本科生知道杨振宁是诺贝尔奖得主，知道他是 Yang-Mills 里的那个 Yang。而为他建一座年鉴，让我看到了教科书略过的东西：他一生中最有分量的一个事实是谁是他的父亲，以及那个父亲在他出生之前，为他铺好了什么。
Published:
When you go looking for a seat in the great theater of the AI era, built from PowerPoints, papers, and anxiety, it helps to look at which play each person on stage is performing.
Published:
当你在这个由PPT、Paper和焦虑构成的AI时代大剧院里找座位时，不如先看看台上的人都在演哪一出戏。
Published:
On a stormy Sunday afternoon in 1988, Edmund Clarke, a Carnegie Mellon professor and Turing Award winner, came to lecture at Fudan University and found the room nearly empty.
Published:
When an AI writing assistant turns a paper’s incoherent rambling into a “well-structured pile of something,” we have to admit: it is basically a cyber laxative.
Published:
当AI写作助手把学术论文里的“狗屁不通”梳理成“结构完整的一大坨”时，我们不得不承认：这简直是一剂赛博开塞露。
Published:
A collection of USTC academic resources covering career planning, library access, and research tools.
Published:
一份涵盖生涯规划、图书馆访问与研究工具的中科大学术资源合集。
Published:
Most physics undergraduates know Chen Ning Yang as a Nobel laureate and the Y in Yang-Mills. Building him a chronicle made me see something the textbooks gloss over: the most consequential single fact about his life is who his father was, and what that father had set up for him before he was born.
Published:
大多数物理本科生知道杨振宁是诺贝尔奖得主，知道他是 Yang-Mills 里的那个 Yang。而为他建一座年鉴，让我看到了教科书略过的东西：他一生中最有分量的一个事实是谁是他的父亲，以及那个父亲在他出生之前，为他铺好了什么。
Published:
On a stormy Sunday afternoon in 1988, Edmund Clarke, a Carnegie Mellon professor and Turing Award winner, came to lecture at Fudan University and found the room nearly empty.
Published:
Today I had an interesting experience collaborating with Claude Code to completely overhaul my personal academic website. As a PhD student in Geological and Earth Sciences at UT Austin, I needed to update my GitHub Pages site with real professional information instead of the placeholder content that had been sitting there.
Published:
今天我经历了一次挺有意思的合作：和 Claude Code 一起，把我的个人学术网站彻底重做了一遍。作为 UT Austin Geological and Earth Sciences 的博士生，我需要把 GitHub Pages 站点从一堆占位符内容，更新成真正能代表我专业背景的信息。
Published:
An AI writing assistant that polishes academic text using DeepSeek-R1 and detects AI-generated content.
Published:
学术写作的严谨性不仅在于数据，更在于表达的精准，而 1AI-polish 通过集成 DeepSeek-R1 的推理能力，为研究者提供了一套集文本润色与 AI 检测于一体的深度协作系统，旨在让复杂的科研思想以更专业、更透明的方式呈现。
Published:
Most academic climate scientists will tell you they care about policy and then publish a paper that no policymaker is ever going to read. Geeta Persad spent four years working at the Union of Concerned Scientists translating climate models for water managers, and she came back to academia knowing exactly what the gap looks like.
Published:
大多数学术气候科学家会告诉你他们在乎政策，然后发表一篇任何决策者都不会读的论文。Geeta Persad 在忧思科学家联盟（Union of Concerned Scientists）花了四年，为水资源管理者翻译气候模型，然后带着对那个差距的清醒认识回到了学术界。
Published:
The sense of agency that took me more than twenty years to grow, these seventeen and eighteen year old students got far more easily, just by being at Stanford, being in America, or maybe going to high school in Shanghai first.
Published:
At today’s Silicon Valley alumni “Ask a Question” AI meet-up, what I wanted to talk about wasn’t how cool Claude Code is, but how it actually re-slices real work.
Published:
今天在硅谷校友会的「师问」AI 交流活动上，我想讲的不是 Claude Code 多酷，而是它到底怎么把真实工作重新拆开。
Published:
Why AI agents lack temporal awareness, and how injecting timestamps transforms them from stateless automata into coherent entities.
Published:
为什么AI Agent缺乏时间意识，以及注入时间戳如何将它们从无状态自动机转变为连贯的实体。
Published:
The signal truly worth heeding is not that AI companies are growing too slowly, but that they are growing fast while cash-flow pressure keeps mounting.
Published:
真正值得警惕的信号不是 AI 公司增长太慢，而是它们增长得很快，现金流压力却仍在放大。
Published:
If you’ve been watching the business map of the AI track lately, you can’t help but be shaken by that trend feverishly racing from “human-machine collaboration” toward “multi-agent autonomy.”
Published:
如果你最近在看AI赛道的商业图谱，你一定会被那个从「人机协作」向「多智能体自治」疯狂演进的趋势给震撼到。
alias rm=trash Cannot Stop an AI Agent from rm -rf Published:
I asked Claude Code a narrow question: can I protect this machine from Codex accidentally deleting files forever, just by aliasing rm to trash? The honest answer turned out to be no, for a reason that is not obvious until you actually test it, and the fix ended up being a five-layer setup rather than a one-liner.
alias rm=trash 拦不住 AI agent 的 rm -rf Published:
我问了 Claude Code 一个很具体的问题：能不能只靠把 rm alias 成 trash，来防止 Codex 意外把文件永久删掉。真实的答案是不能，原因不实测根本看不出来，最后落地的也不是一行配置，而是五层防护。
Published:
Parameterizing a large Fortran climate model by hand is slow, error-prone, and hard to validate. Noah-Agent asks whether a team of specialized AI agents can do it instead.
Published:
手动为大型 Fortran 气候模型配置参数，速度慢、容易出错，也难以验证。Noah-Agent 要检验的是，一组各有所长的 AI 智能体能否协作完成这项工作。
Published:
The Tmux Orchestrator enables Claude agents to work autonomously, schedule their own check-ins, and coordinate across multiple projects without human intervention - a project I explored and learned a lot from.
Published:
Tmux Orchestrator 让 Claude agents 能够自主工作、自己安排 check-in，并在多个项目之间协同推进——这是一个我探索过、也从中学到很多的项目。
Published:
If I reopened the 2026 folder in 2030, I would check one thing first: did the work completed with AI change a real problem and leave evidence that others could verify and build upon?
Published:
假设在 2030 年重新打开 2026 年的文件夹，我会先查一件事：今天借助 AI 完成的工作，是否改变了一个现实问题，并留下别人能够复验和继承的证据。
Published:
I just remembered something sudden on the train: why are almost all the links on the internet that extremely saturated shade of blue?
Published:
我刚刚在火车上突然想起一件事：为什么互联网上的链接，几乎清一色都是那种饱和度极高的蓝色？
Published:
Recently I’ve sometimes felt confused by how far AI has come, a bit lost and anxious, unsure what I should do, and then, because of my identity and my path, wondering what fallbacks or better options exist.
Published:
最近虽然有的时候也会因为 AI 的发展程度感到非常困惑，感到有一点迷茫、有些焦虑，不知道自己该干啥，然后又因为自己的身份和路径问题，在想有什么样的退路或者更好的方案。
Published:
After two years of “All-in AI”, my deep-seated belief in the “10,000-hour talent moat” was ruthlessly shattered by the rise of Agentic workflows.
Published:
在“All-in AI”两年后，我曾经深信的“10,000小时人才护城河”理论，在Agentic工作流的冲击下被无情地打碎了一半。
Published:
Many Chinese teams building embodied AI or AI applications hit the same real-world problem: the domestic models aren’t enough, and they want to call the APIs of foreign models like OpenAI and Anthropic directly. That’s why a swarm of “relay,” “recharge,” and “one OpenAI-compatible interface for you” businesses has sprung up.
Published:
很多做具身智能、做 AI 应用的中国团队，都会遇到同一个现实问题：国内的模型不够用，想直接调 OpenAI、Anthropic 这些海外模型的 API。于是市面上冒出一堆”中转”“代充”“一个 OpenAI 兼容接口给你用”的生意。
Published:
I am deeply honored to join my colleagues in contributing to the newly published paper, “On the Ethics of Generative GeoAI: Explainability, Bias, Hallucination, Accountability, Privacy, and Trust,” specifically focusing on the section regarding GeoAI Trust (Section 8).
Published:
我非常荣幸能够参与到这篇新发表的论文 《On the Ethics of Generative GeoAI: Explainability, Bias, Hallucination, Accountability, Privacy, and Trust》 的合作中，并作为共同作者之一，贡献了关于 GeoAI信任（GeoAI Trust） 的章节（第八章）。
Published:
In the future world, everyone will have their own AI agent assistant.
Published:
在未来的世界，每个人都会有自己的 AI agent 助理。
Published:
Seizing the gateway to the human mind in the AI era is, at bottom, a secret war over token throughput power.
Published:
抢占 AI 时代的人类心智入口，说到底是一场关于 token 吞吐权力的隐秘战争。
Published:
Earlier today I published a piece called “Four Ego Mistakes I Made as a 22-Year-Old Founder.” Then I ran it through Gemini 2.5 Pro as an independent reviewer. Gemini’s verdict was that the essay is itself an ego move. I think Gemini is mostly right.
Published:
今天早些时候我发了一篇叫《22 岁 founder 的四个 ego 错误》。然后我把这篇喂给了 Gemini 2.5 Pro 当独立审稿人。Gemini 的判决是：这篇文章本身就是一个 ego 动作。我觉得 Gemini 大体是对的。
Published:
Call it a (10^5) gap if you like, once you put semantic entropy and effective training samples on the board, the difference between internet-scale text and embodied robotics data stops sounding like rhetorical inflation and starts sounding almost conservative. You just have to separate raw physics bits from semantics a learner can actually use.
Published:
「十万倍」这个说法乍听像聊天里随手甩出去的量级，但如果把它放回语义信息熵和有效训练样本这两把尺子下面，它非但不夸张，反而可能是保守的。关键是你得把「原始物理比特」和「带语义的、可被学习的数据」分开看。
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
Top physics researchers aren’t blind to what AI is doing in the application layer. They’re locked by their own evaluation system and can’t afford to look.
Published:
顶尖的 physics 老年研究者并非看不见应用层的繁荣，而是被自己的评价体系与资源锁定。
Published:
In an era where AI can write code in seconds, I just learned the hard way that blindly moving fast is actually slowing me down.
Published:
在这个闭着眼睛就能让 AI 写代码的时代，我今天结结实实地挨了一锤。
Published:
The scary part is not that AI agents fail on Earth system model code, but that they can fail while looking almost right.
Published:
最可怕的部分不在于 AI 智能体在处理地球系统模型代码时失败了，而是它们可能在看似正确的代码中失败。
Published:
Contribute just one research skill and you can become a coauthor of a Nature paper.
Published:
只要你贡献一点科研绝活，就能成为 Nature 文章的合著者。
Published:
Just by running a single command while using AI for research, you can turn your tacit knowledge into a Nature co-authorship.
Published:
只要你在用AI辅助科研，敲一行命令就能把你脑子里的「隐性经验」变成能发Nature的资本。
Published:
A PhD is hard. A PhD in 2026 — with the field moving faster than your committee can read — is a different kind of hard. AI PhD Survival Guide is a handbook for surviving and finishing an AI or ML PhD without burning out and without falling behind.
Published:
读博很难。2026 年读博——当领域进展快到你的委员会都读不过来时——是另一种难。AI PhD Survival Guide 是一本手册，教你如何在 AI 或机器学习博士研究中活下来并毕业，既不会燃烧殆尽，也不至于掉队。
Published:
Virtual cell models powered by neuromorphic computing and gene language models represent one of the most capital-intensive and scientifically ambitious convergences in biotech AI. If you’re evaluating this space as a VC intern, you need to understand three core components: what these systems actually do, why the market is moving now, and where the investable opportunities lie.
Published:
由神经形态计算和基因语言模型驱动的虚拟细胞模型，代表了生物技术 AI 领域中资本最密集、科学最雄心勃勃的融合方向之一。如果你正在以 VC 实习生的身份评估这个领域，你需要理解三个核心组成部分：这些系统实际上做什么、为什么市场现在开始行动，以及可投资的机会在哪里。
Published:
What Happens When You Just Give People Money? Sam Altman’s UBI Experiment
Published:
直接发钱会怎样？奥特曼的全民基本收入实验
Published:
Why AI agents lack temporal awareness, and how injecting timestamps transforms them from stateless automata into coherent entities.
Published:
为什么AI Agent缺乏时间意识，以及注入时间戳如何将它们从无状态自动机转变为连贯的实体。
Published:
Research, stock trading, and startups all share a common pattern: success comes from combining a defensible core competency with the ability to ride trending waves. You don’t need exhaustive preparation anymore. You need methodology and the ability to produce content when it matters. When the right moment arrives, you strike.
Published:
科研、股市还有创业，这些所有需要展示并能得到结果的东西，本质上都是”主业/具有护城河的本行加上热点”。所以需要做好准备,或者说不一定非要进行那种极其周全的准备，只要掌握一些方法论就可以。利用 AI 在关键时刻能够产出内容，一旦关键热点到来，马上就可以抓住。
Published:
Why next token prediction isn’t just “stochastic parroting”—it’s solving geodesics on the manifold of human knowledge.
Published:
为什么Next Token Prediction不只是”随机鹦鹉”——它在求解人类知识流形上的测地线。
Published:
When continuous functions meet discontinuous truth: Understanding LLM hallucinations through the lens of Fourier analysis.
Published:
当连续函数遇上离散真理：从傅里叶分析的视角理解LLM幻觉。
Published:
When Large Language Models (LLMs) hallucinate, we often treat it as an engineering bug to be fixed. But what if hallucinations are not a flaw, but a mathematical inevitability? A recent interdisciplinary discussion revealed a profound connection between the Gibbs phenomenon in Fourier analysis and the fundamental limitations of neural networks.
Published:
当大语言模型（LLM）产生幻觉时，我们通常将其视为需要修复的工程缺陷。但如果幻觉不是缺陷，而是数学上的必然结果呢？最近一场跨学科讨论揭示了傅里叶分析中的吉布斯现象与神经网络根本局限性之间的深刻联系。
Published:
How photographing a black hole and training GPT-4 are fundamentally the same process: extracting coherent structure from impossibly sparse frequency-domain measurements.
Published:
拍摄黑洞照片和训练GPT-4在根本上是同一个过程：从极度稀疏的频域测量中提取相干结构。
Published:
The promise of AI-automated science is intoxicating: imagine machines that can generate hypotheses, design experiments, and publish papers while we sleep.
Published:
AI 自动化科研的愿景令人陶醉：想象机器能在我们睡觉时生成假设、设计实验、发表论文。然而，尽管关于“AI 科学家”的新闻铺天盖地，我们正在撞上一堵根本性的墙。问题不在于算力或数据集规模，而是某种更深刻的东西，根植于科学发现的本质和信息论之中。
Published:
建立在 19 世纪学徒制模型上的博士培养体系，几乎在所有关键经验指标上都已经出现了系统性失灵。
Published:
I’ve been discussing this question with a friend lately. He wants to use AI for matching in the dating market; I think with a small sample size this is perfectly feasible, you don’t even need to build any search/ads/rec system at all. Just ask the large language model directly, toss in a few users’ profiles, and have it rank them; the whole process is very simple.
Published:
我和朋友最近在讨论这个问题。朋友说想拿 AI 来做婚恋市场的匹配，我觉得在样本量小的情况下，这完全可行——甚至根本不需要搭任何搜广推系统。你直接问大语言模型，把几个用户的简历丢进去，让它排序就好了，整个过程非常简单。
Published:
In the fall of 2015, two things happened simultaneously.
Published:
2015 年秋天，有两件事同时发生了。
Published:
The signal truly worth heeding is not that AI companies are growing too slowly, but that they are growing fast while cash-flow pressure keeps mounting.
Published:
真正值得警惕的信号不是 AI 公司增长太慢，而是它们增长得很快，现金流压力却仍在放大。
Published:
To use OpenClaw or not to use OpenClaw, that is the question.
Published:
用不用 OpenClaw？这是个问题
Published:
Should Silicon Valley engineers use OpenClaw?
Published:
硅谷大厂的工程师究竟需不需要用 OpenClaw？
Published:
Every knowledge worker is now competing with an LLM for some part of their job. AI Survival Guide is a handbook for figuring out which parts, what to do about it, and how to come out the other side better at your craft instead of replaced by it.
Published:
如今，每一位知识工作者都在用某项工作的某些部分与 LLM 竞争。AI Survival Guide 是一本手册，帮你弄清是哪些部分、该怎么应对，以及如何带着更强的技艺从这场竞争的另一头走出来，而不是被它取代。
Published:
If you’ve been watching the business map of the AI track lately, you can’t help but be shaken by that trend feverishly racing from “human-machine collaboration” toward “multi-agent autonomy.”
Published:
如果你最近在看AI赛道的商业图谱，你一定会被那个从「人机协作」向「多智能体自治」疯狂演进的趋势给震撼到。
Published:
Large model vendors spend rocket-launch money raising a legion of geeks who only want to buy a two-dollar firecracker just to hear it pop.
Published:
大模型厂商用造火箭的成本，培养出了一群只想买两块钱鞭炮听个响的极客。
Published:
Sometimes a piece of old science fiction stays in your head not because of the gadgets, but because it got the emotional structure of the future right.
Published:
有些旧科幻会一直留在脑子里，不只是因为它写了什么炫酷科技，而是因为它提前抓住了未来的情感结构。
Published:
Every generation has its own way of managing work: paper notebooks, SaaS task managers, and now programmable agentic workflows powered by tools like OpenClaw heartbeat.
Published:
每一代人都有自己管理任务的方式：最早是纸和笔，后来是 SaaS 任务管理工具，现在则开始进入像 OpenClaw heartbeat 这样可程序化、可持续运行的 agentic workflow 时代。
Published:
For technical people, a strong pull request on a fast-growing open-source repo can become better inbound marketing than another generic post about AI.
Published:
对技术人来说，在一个高速增长的开源仓库里做出高质量 PR，往往比再发一篇泛泛而谈的 AI 观点帖更像真正有效的 inbound marketing。
Published:
Early-career growth usually comes down to a few repeatable layers: alignment, delivery, visibility, communication, leadership, and trust.
Published:
职业早期的成长，通常可以拆成几个可重复的层次：方向对齐、稳定交付、建立可见度、提升沟通、展现领导力，以及赢得信任。
Published:
Lay out every hyped multiAgent product on a table in front of Steve Jobs and he would probably squint, frown, and ask one question, why on earth are you showing the user any of this.
Published:
把现在市面上所有被吹得天花乱坠的multiAgent产品摆在乔布斯面前，他大概率会皱着眉头反问一句，你为什么要让用户看到这些东西。
Published:
If you browse arXiv regularly, you will be struck by the AI papers whose author lists are so long you have to scroll through multiple pages.
Published:
如果你最近经常逛 arxiv，你一定会被那些作者名单长到需要翻页的 AI 论文给震撼到。
Published:
To be totally honest with you, you don’t need to understand the underlying architecture of large language models to get your name on a top-tier AI paper right now, you just need to be a true expert in your own field.
Published:
坦率的讲，你不需要懂任何大模型底层的原理，只要你是一个真的懂行的领域专家，你现在就能在顶级 AI 论文上署名。
Published:
Five YC rebels: some were expelled, some were welcomed back. What happened next?
Published:
五位 YC 叛逆者：有人被驱逐、有人被重新接纳，后来又发生了什么？
Published:
This is the worst startup product I have seen.
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
Published:
At the end of AI’s compute road lies the ultimate solution to energy.
Published:
引言：AI 的算力尽头，是能源的终极解法。
Published:
I have come to despise the word “learning.”
Published:
If you’re serious about understanding LLMs and AI engineering, Simon Willison’s Newsletter is one of the best resources out there. With over 38,000 subscribers, it provides detailed analysis on AI, LLMs, web engineering, open source, data science, and Python.
Published:
如果你认真想要理解 LLM 和 AI 工程，Simon Willison 的 Newsletter 是最好的资源之一。拥有超过 38,000 名订阅者，它提供关于 AI、LLM、Web 工程、开源、数据科学和 Python 的深度分析。
Published:
Every product has its own pros and cons. Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
每一家都有每一家的优点和缺点。Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
In financial markets, we are forever hunting for the kind of “choke point” that everyone calls a stranglehold.
Published:
在金融市场里，我们永远在寻找那种被称为「卡脖子」的咽喉要道。
Published:
Parameterizing a large Fortran climate model by hand is slow, error-prone, and hard to validate. Noah-Agent asks whether a team of specialized AI agents can do it instead.
Published:
手动为大型 Fortran 气候模型配置参数，速度慢、容易出错，也难以验证。Noah-Agent 要检验的是，一组各有所长的 AI 智能体能否协作完成这项工作。
Published:
A hand-controlled 3D visualization of your AI conversation history - fly through ChatGPT conversations and explore artificial intelligence concepts. Try Live Demo
Published:
一个支持手势控制的 3D 可视化项目，把你的 AI 对话历史变成可以飞行探索的星系。你可以在 ChatGPT 对话之间穿梭，也可以把抽象的人工智能概念变成一个能亲手操作的可视化空间。Try Live Demo
Published:
Most crypto dashboards either drown you in numbers or hide the signal behind a paywall. Crypto Dashboard is my attempt at a clean, AI-augmented view of the market that runs entirely in your browser. Try Live Demo
Published:
大多数加密货币仪表盘要么把你淹没在数字里，要么把信号藏在付费墙后面。Crypto Dashboard 是我对一种简洁、由 AI 增强的市场视图的尝试，而且它完全运行在你的浏览器里。试试在线演示
Published:
AlphaEarthHack is the project our team built for the UT Austin Geoscience Hackathon. The goal: see how far we could push AI on Earth system data inside a single weekend. Try Live Demo
Published:
AlphaEarthHack 是我们团队为德州大学奥斯汀分校地球科学黑客松打造的项目。目标只有一个：看看在一个周末内，我们能在地球系统数据上把 AI 推到多远。试试在线演示
Published:
The Tmux Orchestrator enables Claude agents to work autonomously, schedule their own check-ins, and coordinate across multiple projects without human intervention - a project I explored and learned a lot from.
Published:
Tmux Orchestrator 让 Claude agents 能够自主工作、自己安排 check-in，并在多个项目之间协同推进——这是一个我探索过、也从中学到很多的项目。
Published:
A comprehensive tool built with React, Vite, and Supabase to help individuals plan and navigate their careers in the AI industry.
Published:
这是一个基于 React、Vite 和 Supabase 构建的综合工具，帮助个人规划并导航自己在 AI 行业中的职业发展路径。
Published:
An AI writing assistant that polishes academic text using DeepSeek-R1 and detects AI-generated content.
Published:
学术写作的严谨性不仅在于数据，更在于表达的精准，而 1AI-polish 通过集成 DeepSeek-R1 的推理能力，为研究者提供了一套集文本润色与 AI 检测于一体的深度协作系统，旨在让复杂的科研思想以更专业、更透明的方式呈现。
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
Published:
Yuxuan and I have been discussing AI agent ideas for a while. Yesterday, we finally decided: we’re building Atomize, a task-breaking agent system evolved from Goblin Tools.
Published:
我和 Yuxuan 讨论 AI Agent 的想法已经有一段时间了。昨天，我们终于做出了决定：我们要做 Atomize，一个从 Goblin Tools 演化而来的任务分解 Agent 系统。
Published:
The opportunity is not the news itself. The opportunity is the friction created by the news.
Published:
新闻本身不是机会。新闻制造的摩擦才是机会。
Published:
Decades of Earth-system modeling judgment live in PDFs, mailing lists, and senior researchers’ heads, and none of it is loadable by an AI coding agent. earth-space-ai.org is an attempt to fix that.
Published:
数十年来积累的地球系统建模判断散落在 PDF、邮件列表和资深研究者的头脑里，AI 编程智能体无法直接加载这些知识。earth-space-ai.org 试图解决这个问题。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
The University of Texas at Austin’s Office of Graduate and Postdoctoral Studies featured my work on LinkedIn on August 26, 2026.
Published:
德克萨斯大学奥斯汀分校研究生与博士后事务办公室于 2026 年 8 月 26 日在 LinkedIn 上专题介绍了我的工作。
Published:
Port Aransas South Jetty published an Institute Insights feature on May 28, 2026 about how our University of Texas team is using satellite remote sensing and AI to monitor Texas coastal water quality.
Published:
阿兰瑟斯港南防波堤 于 2026 年 5 月 28 日发表了一篇 Institute Insights 专题，讲的是我们德克萨斯大学团队如何用卫星遥感与 AI 监测德克萨斯沿海水质。
Published:
Early-career growth usually comes down to a few repeatable layers: alignment, delivery, visibility, communication, leadership, and trust.
Published:
职业早期的成长，通常可以拆成几个可重复的层次：方向对齐、稳定交付、建立可见度、提升沟通、展现领导力，以及赢得信任。
Published:
The sharpest minds keep moving into AI labs, and it isn’t just for the money.
Published:
这是《颠覆性创新者的思维模式》的五个月后续研究。
Published:
A bit of history from a casual chat: in the US, science students used to get bigger stipends than business, law or medicine students, because science students were poor.
Published:
Today I had an interesting experience collaborating with Claude Code to completely overhaul my personal academic website. As a PhD student in Geological and Earth Sciences at UT Austin, I needed to update my GitHub Pages site with real professional information instead of the placeholder content that had been sitting there.
Published:
今天我经历了一次挺有意思的合作：和 Claude Code 一起，把我的个人学术网站彻底重做了一遍。作为 UT Austin Geological and Earth Sciences 的博士生，我需要把 GitHub Pages 站点从一堆占位符内容，更新成真正能代表我专业背景的信息。
Published:
The promise of AI-automated science is intoxicating: imagine machines that can generate hypotheses, design experiments, and publish papers while we sleep.
Published:
AI 自动化科研的愿景令人陶醉：想象机器能在我们睡觉时生成假设、设计实验、发表论文。然而，尽管关于“AI 科学家”的新闻铺天盖地，我们正在撞上一堵根本性的墙。问题不在于算力或数据集规模，而是某种更深刻的东西，根植于科学发现的本质和信息论之中。
Published:
Call it a (10^5) gap if you like, once you put semantic entropy and effective training samples on the board, the difference between internet-scale text and embodied robotics data stops sounding like rhetorical inflation and starts sounding almost conservative. You just have to separate raw physics bits from semantics a learner can actually use.
Published:
「十万倍」这个说法乍听像聊天里随手甩出去的量级，但如果把它放回语义信息熵和有效训练样本这两把尺子下面，它非但不夸张，反而可能是保守的。关键是你得把「原始物理比特」和「带语义的、可被学习的数据」分开看。
Published:
Exchanging money for time is a privilege I rarely indulge in, but a surprise membership benefit that cleared airport security in ten minutes changed my perspective on friction and market regulation. While a Clear Plus membership normally costs over a hundred dollars annually, obtaining it for free through an Uber membership allowed me to experience a level of efficiency that money can’t always buy—at least not without a well-regulated system behind it.
Published:
用金钱换取时间是一种我很少尝试的“特权”，但这次在奥斯汀机场仅用十分钟便完成安检的经历，让我对效率与市场调节有了新的思考。这份特权并非我主动购买，而是通过 Uber 年费会员赠送的 Clear Plus 获得的——原本需要每年支付一百多美金的服务，在免除门槛后，带给我一种金钱也未必能随时买到的流畅体验。
Published:
I ended up in the wrong NEXUS line at Vancouver International Airport. By the rules, the mistake was mine. By design, it was an almost predictable wrong turn.
Published:
Beyond the wayfinding problem that led me into the wrong NEXUS line, my connection at Vancouver International Airport left me with several observations unrelated to signs. I noticed an older-looking mix of people, a departures area that offered almost nowhere to work, and a visible contrast between people in Vancouver and Texas. This post records those observations. The full account of the NEXUS incident is in How I “Reasonably” Ended Up in the Wrong NEXUS Line at YVR.
Published:
alphaXiv hosts a discussion page for our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation, with the abstract, author list, BibTeX citation, and open discussion.
Published:
alphaXiv 为我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 开设了讨论页，含摘要、作者名单、BibTeX 引用与开放讨论区。
Published:
On February 1 this year, I wrote down a thought about one question: what kind of company is the greatest company in the world?
Published:
今年 2 月 1 日，我写过一段话，想的是一个问题，什么样的公司是世界上最伟大的公司。
Published:
While doing career planning for my girlfriend, I generated this report. Reading it as I went, I found it rather surreal.
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
While doing career planning for my girlfriend, I generated this report. Reading it as I went, I found it rather surreal.
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
At the Metropolitan Museum of Art in New York, I walked into a real ancient Egyptian tomb, not a replica or a model, but a tomb that stood in the Egyptian desert more than four thousand years ago and was moved whole to Manhattan.
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
Published:
I have a theory that the way our bodies are built is adapted to the ancient world.
Published:
Every product has its own pros and cons. Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
每一家都有每一家的优点和缺点。Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
Published:
If all your practice comes from school exams, you will never get that many chances to practice.
Published:
Day fifty of Benchmark Radar. The paper links point at arXiv everywhere, the README asks readers to upvote on Hugging Face, and a dated laurel badge marks the #1 Paper of the Day on the dashboard. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第五十天。论文链接全项目指向 arXiv，README 请读者去 Hugging Face 投票，仪表盘挂起带日期的每日第一论文桂冠。记分牌：185 颗星，30 个 fork。
Published:
Day forty-six of Benchmark Radar. The technical paper is now on arXiv as arXiv:2609.11115, the v0.11.0 release is published with frozen data and verified hashes, and citation surfaces across the project update to the live arXiv entry. Scoreboard: 185 stars, 30 forks.
Published:
Benchmark Radar 的第四十六天。技术论文正式登上 arXiv，编号 arXiv:2609.11115；v0.11.0 发布携带冻结数据与经验证哈希；全项目引用位置更新指向在线 arXiv 条目。记分牌：185 颗星，30 个 fork。
Published:
How can a generative-AI student society do more than host talks, repost news, and chase the latest model?
Published:
一个生成式 AI 学生社团，怎样才能不止于办讲座、转发新闻和追逐下一款模型？
Published:
The sharpest minds keep moving into AI labs, and it isn’t just for the money.
Published:
这是《颠覆性创新者的思维模式》的五个月后续研究。
Published:
The thing that almost stranded me on this international trip was not a typhoon, but several unrelated systems deciding at the same time that they did not trust me.
Published:
这次国际旅行真正差点把我留在原地的，不是台风，而是几个互不认识的系统同时决定不相信我。
Published:
NASA’s Future Investigators in NASA Earth and Space Science and Technology (FINESST) is one of the most underused fellowships among US graduate students. Most PhD students have never heard of it, and the ones who have often miss the deadline because the proposal expectations aren’t obvious from the call alone. This repo is a curated guide of links, tips, and examples to help you write a competitive FINESST proposal.
Published:
NASA 的未来地球与空间科技研究者（FINESST）项目，是美国研究生中最被低估的奖学金之一。这个仓库是一份精心整理的链接、技巧与示例指南，帮助你写出一份有竞争力的 FINESST 提案。
Published:
Research manuscript repository for meteor speed variations study published on Journal of Geophysical Research (JGR): Space Physics.
Published:
这是一个关于流星速度变化研究的论文仓库，相关稿件已在 Journal of Geophysical Research (JGR)发表。
Published:
Have you ever run into this while using AI?
Published:
你在用 AI 的时候，有没有遇到过这种情况？
Published:
I spent a week looking for a cat, got stood up twice, stood someone up once, and on Halloween an orange cat chose me.
Published:
找猫找了一周，被鸽两次，鸽了别人一次，最后在万圣节这一天，被一只橘猫当场选中。
| Unauthorized Release | Résumé of Ms. Pumpkin Patch |
Published:
Guadalupe Street News Agency, Austin, October 31
Published:
瓜达卢佩街通讯社奥斯汀10月31日电
Published:
Somebody put up a poster. The poster has eight words and a QR code.
Published:
有人贴了一张海报。海报上只有八个英文单词，和一个二维码。
Published:
Tomorrow, Amazon delivers two boxes to my apartment.
Published:
明天，亚马逊要给我家送来两个箱子。
Published:
The thing that almost stranded me on this international trip was not a typhoon, but several unrelated systems deciding at the same time that they did not trust me.
Published:
这次国际旅行真正差点把我留在原地的，不是台风，而是几个互不认识的系统同时决定不相信我。
Published:
What I actually want to do today is pull myself out from under a mountain of repetitive work.
Published:
我今天真正想做的事情，其实是把自己从一座 mountains of 重复性的 work 里面捞出来。
Published:
To use OpenClaw or not to use OpenClaw, that is the question.
Published:
用不用 OpenClaw？这是个问题
Published:
Today I had an interesting experience collaborating with Claude Code to completely overhaul my personal academic website. As a PhD student in Geological and Earth Sciences at UT Austin, I needed to update my GitHub Pages site with real professional information instead of the placeholder content that had been sitting there.
Published:
今天我经历了一次挺有意思的合作：和 Claude Code 一起，把我的个人学术网站彻底重做了一遍。作为 UT Austin Geological and Earth Sciences 的博士生，我需要把 GitHub Pages 站点从一堆占位符内容，更新成真正能代表我专业背景的信息。
Published:
The Tmux Orchestrator enables Claude agents to work autonomously, schedule their own check-ins, and coordinate across multiple projects without human intervention - a project I explored and learned a lot from.
Published:
Tmux Orchestrator 让 Claude agents 能够自主工作、自己安排 check-in，并在多个项目之间协同推进——这是一个我探索过、也从中学到很多的项目。
Published:
@elonmusk, asking for a friend, has Tesla FSD been tested in weather like this?
Published:
Checking flights today, I suddenly noticed: AI is the airline code for Air India.
Published:
今天看航班时，我突然发现：AI 是 Air India（印度航空）的航空公司代码。
Published:
USTC recognized my project with the 2023 Undergraduate Research Plan Outstanding Award. The project used WACCM-X and observations to study how the 2022 Tonga eruption disturbed the mesosphere and lower thermosphere.
Published:
我的项目获得中国科学技术大学2023 年大学生研究计划校级优秀奖。项目利用 WACCM-X 和观测数据，研究 2022 年汤加火山喷发对中间层与低热层区域的扰动。
Published:
USTC awarded my undergraduate thesis a 2024 Outstanding Graduation Thesis Award. The thesis studied the 2022 Tonga volcanic eruption’s disturbance of the mesosphere and lower thermosphere using WACCM-X simulations and observations.
Published:
我的本科毕业论文获得中国科学技术大学2024 年度校级优秀毕业论文奖。论文利用 WACCM-X 模拟和观测数据，研究 2022 年汤加火山喷发对中间层与低热层区域的扰动。
Published:
The USTC Library published the results of the university’s 2020 campus calligraphy competition. My entry received third prize.
Published:
中国科大图书馆发布了 2020 年校园书法比赛结果。我的参赛作品获得三等奖。
Published:
As we walked out of the library, an elderly white woman with silver hair saw us and said “excuse me” three times. She spoke to us with real pride, took out her phone, and said she had seen the two of us taking photos with the frog statue out front. Every Christmas, she told us, that frog gets dressed up in a Christmas outfit, look, just like this! Then she asked, very sincerely, whether we would still be here at Christmas, and how we had even heard of Atherton, because the place is so unremarkable, so quiet, so low-key. We told her we would probably each be heading home by then. She added that her son had been learning Chinese since he was a little kid, and that he speaks it really well.
Published:
If you ask most people where trees in California get their water in a drought, they will say “the soil.” It turns out a huge fraction of it comes from cracks in the bedrock underneath the soil, and Daniella Rempe is the person who put numbers on it.
Published:
如果你问大多数人，加州在干旱时树从哪里取水，他们会说”土壤”。事实证明，很大一部分水其实来自土壤下方岩石裂缝里的基岩，而 Daniella Rempe 就是把数字放到这件事上的人。
Published:
The scary part is not that AI agents fail on Earth system model code, but that they can fail while looking almost right.
Published:
最可怕的部分不在于 AI 智能体在处理地球系统模型代码时失败了，而是它们可能在看似正确的代码中失败。
Published:
I submitted our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation to Hugging Face Daily Papers on September 14, 2026, and readers voted it to the top of that day’s list. As of September 29 it has 171 upvotes, the most of the 26 papers posted that day.
Published:
2026 年 9 月 14 日，我把论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 提交到 Hugging Face Daily Papers，读者把它投上了当天榜首。截至 9 月 29 日，它获得 171 个赞，在当天 26 篇论文中排第一。
Published:
The SciFi account shared Benchmark Radar on X, carrying the paper beyond the eval community into adjacent audiences.
Published:
SciFi 账号在 X 转发 Benchmark Radar，把论文带出评测圈子，送往相邻受众。
Published:
HuggingPapers mentioned Benchmark Radar on X, carrying arXiv:2609.11115 into the daily paper flow.
Published:
HuggingPapers 在 X 提及 Benchmark Radar，把 arXiv:2609.11115 送进每日论文流。
Published:
A community share carried Benchmark Radar further across X, reader to readers.
Published:
一条社区转发把 Benchmark Radar 带往 X 更远处，读者传读者。
Published:
A Moltbook post carried Benchmark Radar onto the agent internet’s front page — the social feed built for agents, by agents.
Published:
一条 Moltbook 帖子把 Benchmark Radar 送上 agent 互联网的首页——那个为 agent 建、由 agent 用的社交 feed。
Published:
Alexy Khrabrov, who runs Strongly Typed AI News, shared its Benchmark Radar lead story on LinkedIn from his own account, carrying it from the news site into professional networks.
Published:
Strongly Typed AI News 的运营者 Alexy Khrabrov 用个人账号在 LinkedIn 转发了 Benchmark Radar 头条，把它从新闻站送进职业网络。
Published:
A launch-coverage share on LinkedIn carried the Benchmark Radar launch story into professional networks.
Published:
一条 LinkedIn 发布报道转发把 Benchmark Radar 发布故事送进职业网络。
Published:
The Isfahan AI Desk (میز هوش مصنوعی استان اصفهان), the province’s AI governance and ecosystem watch, listed our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation (arXiv:2609.11115) in its news radar with a relevance score of 91/100.
Published:
伊斯法罕省 AI 治理与生态观察（میز هوش مصنوعی استان اصفهان）把我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation（arXiv:2609.11115）列入新闻雷达，相关度评分 91/100。
Published:
Strongly Typed AI News, the AI news site Alexy Khrabrov runs at collected.ga, made our paper the lead story of its September 12, 2026 “Typesafe AI Daily” issue. Its crawler checked 49 sources that day and found 12 strong signals across AI infrastructure, funding, research, and developer tools, and it named Benchmark Radar the strongest of them.
Published:
Alexy Khrabrov 在 collected.ga 运营的 AI 新闻站 Strongly Typed AI News 把我们的论文做成 2026 年 9 月 12 日「Typesafe AI Daily」一期的头条。当天爬虫检查 49 个来源，在 AI 基建、融资、研究、开发者工具中找出 12 条强信号，Benchmark Radar 排第一。
Published:
Papers with Code listed our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation (arXiv:2609.11115), with its TL;DR, abstract, code repository, and project page.
Published:
Papers with Code 收录了我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation（arXiv:2609.11115），附 TL;DR、摘要、代码仓库与项目页。
Published:
Moonlight, the AI research-reading companion by Corca, published a literature-review page for our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation.
Published:
AI 论文阅读助手 Moonlight（Corca 出品）为我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 发布了文献解读页。
Published:
Geepity turned our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation into a full interactive reading edition, walking through all nine chapters with live charts and a working BM25F demo.
Published:
Geepity 把我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 做成完整互动解读版，九章走完，配实时图表与可运行的 BM25F 演示。
Published:
DailyAIWire covered the Benchmark Radar launch as a centralized AI evaluation database, with an AI-assisted intelligence report, an explain-like-I’m-five summary, and a visual flow of the system.
Published:
DailyAIWire 以中心化 AI 评测数据库为题报道了 Benchmark Radar 发布，附 AI 辅助情报分析、五岁小孩版解释与系统流程图。
Published:
arXivDaily, the Chinese arXiv digest with per-university pages, lists our paper on its Shanghai Jiao Tong University page, following the SJTU affiliation on the author list.
Published:
中文 arXiv 速递 arXivDaily 在上海交通大学高校专区收录了我们的论文，作者名单里的上交 affiliation 把它带到这里。
Published:
alphaXiv hosts a discussion page for our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation, with the abstract, author list, BibTeX citation, and open discussion.
Published:
alphaXiv 为我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 开设了讨论页，含摘要、作者名单、BibTeX 引用与开放讨论区。
Published:
你好，我是 Koutian。第十五天，雷达第一次学会主动对外说话，不再只是等你来翻。
Published:
你好，我是 Koutian。第十四天，我们把雷达的几个老毛病修好了，也让它更像一个会自己干活的工具。
Published:
第十三天的重点是让仪表盘更好看，并且把每条结论都说得有凭有据。
Published:
雷达现在能订阅了。第十二天，我们加了一个每天更新的 RSS 订阅源。
Published:
第十一天，KW-Bench 的能力分级上线了，社区发布也开始了。
Published:
Your next cloud instance is alive, it needs to be fed every three days, and it will be dead in six months whether you cancel or not.
Published:
Most of biology is averages. You measure a million cells, you get a mean. Eric Greene’s lab figured out how to actually watch one DNA molecule at a time, repair itself, in real time, and that changes what kind of questions biology can ask.
Published:
大部分生物学是平均值。你测一百万个细胞，得到一个均值。Eric Greene 的实验室却想出了怎么一次实时观看一个 DNA 分子自我修复，而这改变了生物学能提出什么样的问题。
Published:
Virtual cell models powered by neuromorphic computing and gene language models represent one of the most capital-intensive and scientifically ambitious convergences in biotech AI. If you’re evaluating this space as a VC intern, you need to understand three core components: what these systems actually do, why the market is moving now, and where the investable opportunities lie.
Published:
The most important measure of work may be the energy it consumes, more than the time it takes.
Published:
衡量工作效率，最重要的指标或许不是投入的时间，而是消耗的能源。
Published:
Most crypto dashboards either drown you in numbers or hide the signal behind a paywall. Crypto Dashboard is my attempt at a clean, AI-augmented view of the market that runs entirely in your browser. Try Live Demo
Published:
大多数加密货币仪表盘要么把你淹没在数字里，要么把信号藏在付费墙后面。Crypto Dashboard 是我对一种简洁、由 AI 增强的市场视图的尝试，而且它完全运行在你的浏览器里。试试在线演示
Published:
When something urgent lands at work and there is no time to prepare, so you stay up all night, that is a black swan.
Published:
Sometimes a piece of old science fiction stays in your head not because of the gadgets, but because it got the emotional structure of the future right.
Published:
有些旧科幻会一直留在脑子里，不只是因为它写了什么炫酷科技，而是因为它提前抓住了未来的情感结构。
Published:
If you’ve been watching the business map of the AI track lately, you can’t help but be shaken by that trend feverishly racing from “human-machine collaboration” toward “multi-agent autonomy.”
Published:
如果你最近在看AI赛道的商业图谱，你一定会被那个从「人机协作」向「多智能体自治」疯狂演进的趋势给震撼到。
Published:
The more I think about it, the more I believe that doing business, commerce and trade, is the most fundamental form of charity.
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
Most people think of summer vacation as time off. I started wondering whether that framing undersells it. Summer Calculator is a small web tool that estimates the full value of your summer — money, learning, relationships, health — not just the days you didn’t work. Try Live Demo
Published:
大多数人把暑假当作休息时间。我开始怀疑这种框架是否低估了它。Summer Calculator 是一款小巧的网页工具，用来估算你暑假的完整价值——金钱、学习、人际关系、健康——而不仅仅是你没工作的那些日子。试试在线演示
Published:
School rankings are a single number standing in for a multi-dimensional decision. School Evaluator is a small web tool that lets you score a school across the dimensions that actually shape your life there. Try Live Demo
Published:
学校排名只是一个用单一数字去代表多维决策的替代品。School Evaluator 是一款小巧的网页工具，让你能在一个真正决定你在那里生活的维度上为学校打分。试试在线演示
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
第十一天，KW-Bench 的能力分级上线了，社区发布也开始了。
Published:
Recently I’ve sometimes felt confused by how far AI has come, a bit lost and anxious, unsure what I should do, and then, because of my identity and my path, wondering what fallbacks or better options exist.
Published:
最近虽然有的时候也会因为 AI 的发展程度感到非常困惑，感到有一点迷茫、有些焦虑，不知道自己该干啥，然后又因为自己的身份和路径问题，在想有什么样的退路或者更好的方案。
Published:
After two years of “All-in AI”, my deep-seated belief in the “10,000-hour talent moat” was ruthlessly shattered by the rise of Agentic workflows.
Published:
在“All-in AI”两年后，我曾经深信的“10,000小时人才护城河”理论，在Agentic工作流的冲击下被无情地打碎了一半。
Published:
Top physics researchers aren’t blind to what AI is doing in the application layer. They’re locked by their own evaluation system and can’t afford to look.
Published:
While doing career planning for my girlfriend, I generated this report. Reading it as I went, I found it rather surreal.
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
While doing career planning for my girlfriend, I generated this report. Reading it as I went, I found it rather surreal.
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
Research, stock trading, and startups all share a common pattern: success comes from combining a defensible core competency with the ability to ride trending waves. You don’t need exhaustive preparation anymore. You need methodology and the ability to produce content when it matters. When the right moment arrives, you strike.
Published:
Every knowledge worker is now competing with an LLM for some part of their job. AI Survival Guide is a handbook for figuring out which parts, what to do about it, and how to come out the other side better at your craft instead of replaced by it.
Published:
如今，每一位知识工作者都在用某项工作的某些部分与 LLM 竞争。AI Survival Guide 是一本手册，帮你弄清是哪些部分、该怎么应对，以及如何带着更强的技艺从这场竞争的另一头走出来，而不是被它取代。
Published:
Every generation has its own way of managing work: paper notebooks, SaaS task managers, and now programmable agentic workflows powered by tools like OpenClaw heartbeat.
Published:
每一代人都有自己管理任务的方式：最早是纸和笔，后来是 SaaS 任务管理工具，现在则开始进入像 OpenClaw heartbeat 这样可程序化、可持续运行的 agentic workflow 时代。
Published:
For technical people, a strong pull request on a fast-growing open-source repo can become better inbound marketing than another generic post about AI.
Published:
对技术人来说，在一个高速增长的开源仓库里做出高质量 PR，往往比再发一篇泛泛而谈的 AI 观点帖更像真正有效的 inbound marketing。
Published:
Early-career growth usually comes down to a few repeatable layers: alignment, delivery, visibility, communication, leadership, and trust.
Published:
职业早期的成长，通常可以拆成几个可重复的层次：方向对齐、稳定交付、建立可见度、提升沟通、展现领导力，以及赢得信任。
Published:
Walking University Avenue in Palo Alto today, I finally saw what many Chinese students get wrong, we treat Silicon Valley as a destination.
Published:
To be totally honest with you, you don’t need to understand the underlying architecture of large language models to get your name on a top-tier AI paper right now, you just need to be a true expert in your own field.
Published:
坦率的讲，你不需要懂任何大模型底层的原理，只要你是一个真的懂行的领域专家，你现在就能在顶级 AI 论文上署名。
Published:
There is no top-tier, universally applicable methodology, except one. Go deep in one industry, build personal influence, build a personal brand.
Published:
Jiayi Weng planned his career at OpenAI around one goal, to get his name onto as many OpenAI blog posts as possible.
Published:
Doing research is gambling. In fact, for people without money, research is gambling, and so is everything else.
Published:
Here is a view I hold now: whatever you do, you look a lot like a speculator in a market.
Published:
Sunk cost should never be a reason for a decision, and that includes your choice of major.
Published:
Why do some people who seem to work less than you succeed earlier? Because success is a question of timing first, and ability second.
Published:
You think you picked your major rationally, but the choice itself is a bet.
Published:
A comprehensive tool built with React, Vite, and Supabase to help individuals plan and navigate their careers in the AI industry.
Published:
这是一个基于 React、Vite 和 Supabase 构建的综合工具，帮助个人规划并导航自己在 AI 行业中的职业发展路径。
Published:
This afternoon I paid 80 dollars at an animal shelter in Austin and brought home a three-year-old orange cat.
Published:
I spent a week looking for a cat, got stood up twice, stood someone up once, and on Halloween an orange cat chose me.
Published:
找猫找了一周，被鸽两次，鸽了别人一次，最后在万圣节这一天，被一只橘猫当场选中。
| Unauthorized Release | Résumé of Ms. Pumpkin Patch |
Published:
Guadalupe Street News Agency, Austin, October 31
Published:
瓜达卢佩街通讯社奥斯汀10月31日电
Published:
Tomorrow, Amazon delivers two boxes to my apartment.
Published:
明天，亚马逊要给我家送来两个箱子。
Published:
The more I think about it, the more I believe that doing business, commerce and trade, is the most fundamental form of charity.
Published:
In October 2023 I asked ChatGPT an over-engineered question: how do you fry rice so that every grain of rice ends up bonded to egg, with no bare rice grains and no isolated clumps of egg sitting off on their own? I saved that conversation as an HTML file, dropped it in my Downloads folder, and did not look at it again for almost three years. Last week I finally turned it into an actual open-source repo, and going back through the original conversation to build it was more interesting than I expected.
Published:
2023 年 10 月，我问了 ChatGPT 一个过度工程化的问题：怎么炒蛋炒饭，才能让每一粒米都裹上蛋，既没有裸露的白米粒，也没有单独抱团、没沾到米的蛋碎？那次对话我存成了一个 HTML 文件，扔进 Downloads 文件夹，然后差不多三年没再打开过。上周我终于把它整理成了一个真正的开源仓库，回头重读那段 2023 年的对话、把它做成代码的过程，比我预想的有意思得多。
Published:
In the fall of 2015, two things happened simultaneously.
Published:
2015 年秋天，有两件事同时发生了。
Published:
What do a 7,169-yuan Chinese animated film and a 483-dollar GameStop short squeeze have in common? More than you would think.
Published:
一部票房 7169 元的国产动画，和一只冲到 483 美元的 GameStop 股票，有什么共同点？比你想象的要多。
Published:
“这b人生过的值不值” — roughly, “was this damn life worth it?” — is a phrase that does a lot of emotional work on the Chinese internet. ZuoGuoHuaDiao (做过划掉) is a small web tool that takes the phrase seriously and tries to answer it with numbers. Try Live Demo
Published:
“这b人生过的值不值”——简单说，就是”这该死的一生过得值不值？”——这句在中国互联网上承载了大量情感的话语。ZuoGuoHuaDiao（做过划掉）是一款小巧的网页工具，它认真对待这句话，并试图用数字来回答它。试试在线演示
Published:
The US is full of small traps that nobody warns you about until after you’ve fallen into one. US Survival Guide is a handbook for Chinese students and newcomers on identifying risk early, knowing your rights, and building a financial defense before you need it. Read the Guide
Published:
美国布满了各式各样的暗坑，在你掉进去之前，没人会提醒你。US Survival Guide 是一本写给中国学生和新移民的手册，教你及早识别风险、了解自己的权利，并在你需要之前先筑好财务防线。阅读指南
Published:
Most physics undergraduates know Chen Ning Yang as a Nobel laureate and the Y in Yang-Mills. Building him a chronicle made me see something the textbooks gloss over: the most consequential single fact about his life is who his father was, and what that father had set up for him before he was born.
Published:
大多数物理本科生知道杨振宁是诺贝尔奖得主，知道他是 Yang-Mills 里的那个 Yang。而为他建一座年鉴，让我看到了教科书略过的东西：他一生中最有分量的一个事实是谁是他的父亲，以及那个父亲在他出生之前，为他铺好了什么。
Published:
Two years later, I am back in this familiar city.
Published:
Just by running a single command while using AI for research, you can turn your tacit knowledge into a Nature co-authorship.
Published:
只要你在用AI辅助科研，敲一行命令就能把你脑子里的「隐性经验」变成能发Nature的资本。
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
alias rm=trash Cannot Stop an AI Agent from rm -rf Published:
I asked Claude Code a narrow question: can I protect this machine from Codex accidentally deleting files forever, just by aliasing rm to trash? The honest answer turned out to be no, for a reason that is not obvious until you actually test it, and the fix ended up being a five-layer setup rather than a one-liner.
alias rm=trash 拦不住 AI agent 的 rm -rf Published:
我问了 Claude Code 一个很具体的问题：能不能只靠把 rm alias 成 trash，来防止 Codex 意外把文件永久删掉。真实的答案是不能，原因不实测根本看不出来，最后落地的也不是一行配置，而是五层防护。
Published:
Should Silicon Valley engineers use OpenClaw?
Published:
硅谷大厂的工程师究竟需不需要用 OpenClaw？
Published:
When you join a lab, you do not just get a research direction. You inherit a 30-year codebase that is currently running inside the U.S. National Water Model and was on the critical path forecasting Hurricane Harvey. That is what working with Zong-Liang Yang at the Jackson School of Geosciences actually looks like.
Published:
加入实验室以后，你会接过一个研究方向，也会继承一套已有 30 年历史的代码库。它目前运行在美国国家水模型中，也曾处在飓风 Harvey 预报工作的关键路径上。这就是在 Jackson School of Geosciences 与 Zong-Liang Yang 一起工作的真实样子。
Published:
The scary part is not that AI agents fail on Earth system model code, but that they can fail while looking almost right.
Published:
最可怕的部分不在于 AI 智能体在处理地球系统模型代码时失败了，而是它们可能在看似正确的代码中失败。
Published:
AlphaEarthHack is the project our team built for the UT Austin Geoscience Hackathon. The goal: see how far we could push AI on Earth system data inside a single weekend. Try Live Demo
Published:
AlphaEarthHack 是我们团队为德州大学奥斯汀分校地球科学黑客松打造的项目。目标只有一个：看看在一个周末内，我们能在地球系统数据上把 AI 推到多远。试试在线演示
Published:
Dr. Zong-Liang Yang’s Land Environment and Atmospheric Dynamics (LEAD) Group at UT-Austin employs satellite remote sensing, earth system modeling, and high-performance computing to advance understanding of Earth system sciences.
Published:
UT Austin 宗良杨博士（Dr. Zong-Liang Yang）领导的陆地环境与大气动力学研究组（LEAD Group）利用卫星遥感、地球系统建模和高性能计算，推动人类对地球系统科学的理解。
Published:
Most academic climate scientists will tell you they care about policy and then publish a paper that no policymaker is ever going to read. Geeta Persad spent four years working at the Union of Concerned Scientists translating climate models for water managers, and she came back to academia knowing exactly what the gap looks like.
Published:
大多数学术气候科学家会告诉你他们在乎政策，然后发表一篇任何决策者都不会读的论文。Geeta Persad 在忧思科学家联盟（Union of Concerned Scientists）花了四年，为水资源管理者翻译气候模型，然后带着对那个差距的清醒认识回到了学术界。
Published:
Most academic climate scientists will tell you they care about policy and then publish a paper that no policymaker is ever going to read. Geeta Persad spent four years working at the Union of Concerned Scientists translating climate models for water managers, and she came back to academia knowing exactly what the gap looks like.
Published:
大多数学术气候科学家会告诉你他们在乎政策，然后发表一篇任何决策者都不会读的论文。Geeta Persad 在忧思科学家联盟（Union of Concerned Scientists）花了四年，为水资源管理者翻译气候模型，然后带着对那个差距的清醒认识回到了学术界。
Published:
Decades of Earth-system modeling judgment live in PDFs, mailing lists, and senior researchers’ heads, and none of it is loadable by an AI coding agent. earth-space-ai.org is an attempt to fix that.
Published:
数十年来积累的地球系统建模判断散落在 PDF、邮件列表和资深研究者的头脑里，AI 编程智能体无法直接加载这些知识。earth-space-ai.org 试图解决这个问题。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
Parameterizing a large Fortran climate model by hand is slow, error-prone, and hard to validate. Noah-Agent asks whether a team of specialized AI agents can do it instead.
Published:
手动为大型 Fortran 气候模型配置参数，速度慢、容易出错，也难以验证。Noah-Agent 要检验的是，一组各有所长的 AI 智能体能否协作完成这项工作。
Published:
The roar of cooling systems in the Texas Advanced Computing Center (TACC) isn’t just noise—it’s the power of supercomputers translating the overwhelming dimensionality of climate data into something students can finally see and understand.
Published:
德州高级计算中心（TACC）里冷却系统的轰鸣不只是噪音——那是超级计算机的力量，正在把气候数据那令人不知所措的多维性，转化成学生最终能看见、能理解的东西。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
Your next cloud instance is alive, it needs to be fed every three days, and it will be dead in six months whether you cancel or not.
alias rm=trash Cannot Stop an AI Agent from rm -rf Published:
I asked Claude Code a narrow question: can I protect this machine from Codex accidentally deleting files forever, just by aliasing rm to trash? The honest answer turned out to be no, for a reason that is not obvious until you actually test it, and the fix ended up being a five-layer setup rather than a one-liner.
alias rm=trash 拦不住 AI agent 的 rm -rf Published:
我问了 Claude Code 一个很具体的问题：能不能只靠把 rm alias 成 trash，来防止 Codex 意外把文件永久删掉。真实的答案是不能，原因不实测根本看不出来，最后落地的也不是一行配置，而是五层防护。
Published:
Going around GitHub opening issues and pull requests is sperm donation. The same mechanism, running as-is.
Published:
在 GitHub 上到处开 issue、提 pull request，就是赛博捐精。同一套机制，原样运行。
Published:
Let me offer a comparison that’s a bit inappropriate but does capture a real structure of open-source collaboration:
Published:
我提一个不太恰当、但确实抓住了开源协作某种结构的比喻：
Published:
In the future world, everyone will have their own AI agent assistant.
Published:
在未来的世界，每个人都会有自己的 AI agent 助理。
Published:
If you browse arXiv regularly, you will be struck by the AI papers whose author lists are so long you have to scroll through multiple pages.
Published:
如果你最近经常逛 arxiv，你一定会被那些作者名单长到需要翻页的 AI 论文给震撼到。
Published:
Here is a view I hold now: whatever you do, you look a lot like a speculator in a market.
Published:
School rankings are a single number standing in for a multi-dimensional decision. School Evaluator is a small web tool that lets you score a school across the dimensions that actually shape your life there. Try Live Demo
Published:
学校排名只是一个用单一数字去代表多维决策的替代品。School Evaluator 是一款小巧的网页工具，让你能在一个真正决定你在那里生活的维度上为学校打分。试试在线演示
Published:
A widely shared post describes a brainstorming meeting at a robotics startup in Boston. A slide said “No stupid questions.” A Chinese employee interpreted it as “Do not ask stupid questions,” while American colleagues explained it as “Ask freely; no question will be judged stupid.” A related post claims that “Say it again?” is neutral when someone is not heard, whereas “What did you say?” is hostile.
Published:
一张流传截图讲了这样一件事：一家波士顿机器人创业公司开头脑风暴会，幻灯片写着 “No stupid questions.” 一位中国员工把它理解为“不要问蠢问题”，美国同事却说它的意思是“什么都可以问，没有问题会被当成蠢问题”。另一组讨论又声称：没听清时说 “Say it again?” 很中性，而 “What did you say?” 很不友好。
Published:
When you send an urgent email asking for help, “quick question” gets a reply more easily than “emergency.”
Published:
The USTC Student Club Association recorded the founding of the Xingyun Poetry Club, which I helped start. The club created a home for poetry readings, writing, and student-run cultural events; its activities and posts later reached more than 100,000 views across platforms.
Published:
中国科学技术大学社团管指委记录了星云诗社的创办，我是发起人之一。诗社为诗歌朗读、写作和学生自办文化活动提供了一个长期空间，相关活动和内容后来在各平台累计获得十万余次浏览。
Published:
瀚海奇点 covered USTC’s 2021 Science Fiction Night. I took part in the event, which brought students together around science fiction, ideas, and creative work.
Published:
“瀚海奇点”报道了中国科大 2021 年科幻之夜。我参与了这场活动，与同学们一起交流科幻、想法和创作。
Published:
Starting a petition, founding a club, and joining a research effort someone else organized look like three different things. They all run into the same question: who decides, who puts in the time, and when something goes wrong, who bears the consequences?
Published:
发起请愿、成立社团、参加别人组织的科研活动，看上去是三件不同的事。它们都绕不开同一个问题：谁作决定，谁投入时间，事情出了问题以后，谁承担后果？
Published:
How can a generative-AI student society do more than host talks, repost news, and chase the latest model?
Published:
一个生成式 AI 学生社团，怎样才能不止于办讲座、转发新闻和追逐下一款模型？
Published:
第十一天，KW-Bench 的能力分级上线了，社区发布也开始了。
Published:
Rereading my old journals today, from nuclear fusion in physics to the games humans play, I keep finding the same thing. Whether it is a startup or a whole industry, what matters is a force that rises on its own and then keeps itself going.
Published:
Secretly recording a customer interview can turn a useful research habit into a privacy problem in one click.
Published:
偷偷录一次客户访谈，可能把一个正常的产品研究流程变成隐私和合规事故。
Published:
Many Chinese teams building embodied AI or AI applications hit the same real-world problem: the domestic models aren’t enough, and they want to call the APIs of foreign models like OpenAI and Anthropic directly. That’s why a swarm of “relay,” “recharge,” and “one OpenAI-compatible interface for you” businesses has sprung up.
Published:
很多做具身智能、做 AI 应用的中国团队，都会遇到同一个现实问题：国内的模型不够用，想直接调 OpenAI、Anthropic 这些海外模型的 API。于是市面上冒出一堆”中转”“代充”“一个 OpenAI 兼容接口给你用”的生意。
Published:
People love saying that ten years to a billion-dollar valuation is the normal pace for a unicorn, neither fast nor slow. But once you actually crunch the compound numbers, you realize what hides behind that word “normal” is a beast that doubles in value every single year.
Published:
For a long time I thought adding thousands of people on LinkedIn by hand, just for fun, was all show and no use.
Published:
Don’t expect your first video to go viral, and don’t expect your first sales trip to land ten million dollars, but you can ask what happens after the hundredth try.
Published:
The talk that stayed with me at the Texas Symposium on Machine Learning, Responsible AI & Robotics 2026 was about extending language models, not just making them bigger.
Published:
In my first post on this I described sending a patch through GitGitGadget and watching it land on the mailing list as v1. That post ended with the patch waiting for review. It has since merged. Now that I have seen the full lifecycle, from typo to master, here is what actually mattered.
Published:
在第一篇文章里，我写了怎么通过 GitGitGadget 把补丁发出去，它作为 v1 落在了邮件列表上。那篇文章结束在等待 review 的状态。现在它已经合并了。看完从一个拼写错误到进入 master 的完整过程，下面是我觉得真正重要的几点。
Published:
The patch I sent to the GitGitGadget doorstep ten days ago has been merged into master.
Published:
十天前我送到 GitGitGadget 门口的那个补丁，现在已经合并进 master 了。
Published:
Today I got my first Git patch onto the mailing list.
Published:
今天我差一点，就把自己的第一个 Git 项目补丁送进邮件列表了。
Published:
One hundred twenty-six thousand eight hundred and eighty-seven dollars is housing-level money in Austin; in UT Austin’s new business-school building, it buys a single parking spot.
Published:
十二万六千八百八十七美元，在 Austin 已经是住房级别的钱；在 UT Austin 的新商学院大楼里，它只对应一个停车位。
Published:
A long time ago, I came across a book called Chase, Chance, and Creativity: The Lucky Art of Novelty（《追逐、机遇和创造力：新奇的幸运艺术》）. I no longer remember everything in it, but its central idea stayed somewhere in the back of my mind: luck and creativity are not entirely random. Chance may arrive unexpectedly, but we can still choose how often we encounter it and whether we are ready to recognize it.
Published:
很久以前，我偶然看到一本书，叫 Chase, Chance, and Creativity: The Lucky Art of Novelty（《追逐、机遇和创造力：新奇的幸运艺术》）。我已经不记得书里的全部内容了，但它的核心想法一直留在我脑海中的某个角落：运气和创造力并不完全是随机的。机遇也许会意外到来，但我们仍然可以选择自己遇见它的频率，以及当它出现时，我们是否已经准备好认出它。
Published:
A receipt in Singapore showed S$5.50, while a U.S. credit-card account displayed only US$4.27. Another purchase of about S$48 appeared as roughly US$38. A bank-card transit ride also failed to appear immediately as pending.
Published:
在新加坡消费时，一张收据写着 S$5.50，但美国信用卡账户只显示 US$4.27。另一笔约 S$48 的消费，则显示为约 US$38。与此同时，刷银行卡乘坐公共交通后，交易也没有立刻出现在 pending 中。
Published:
If you ask most people where trees in California get their water in a drought, they will say “the soil.” It turns out a huge fraction of it comes from cracks in the bedrock underneath the soil, and Daniella Rempe is the person who put numbers on it.
Published:
如果你问大多数人，加州在干旱时树从哪里取水，他们会说”土壤”。事实证明，很大一部分水其实来自土壤下方岩石裂缝里的基岩，而 Daniella Rempe 就是把数字放到这件事上的人。
Published:
The US is full of small traps that nobody warns you about until after you’ve fallen into one. US Survival Guide is a handbook for Chinese students and newcomers on identifying risk early, knowing your rights, and building a financial defense before you need it. Read the Guide
Published:
美国布满了各式各样的暗坑，在你掉进去之前，没人会提醒你。US Survival Guide 是一本写给中国学生和新移民的手册，教你及早识别风险、了解自己的权利，并在你需要之前先筑好财务防线。阅读指南
Published:
Most crypto dashboards either drown you in numbers or hide the signal behind a paywall. Crypto Dashboard is my attempt at a clean, AI-augmented view of the market that runs entirely in your browser. Try Live Demo
Published:
大多数加密货币仪表盘要么把你淹没在数字里，要么把信号藏在付费墙后面。Crypto Dashboard 是我对一种简洁、由 AI 增强的市场视图的尝试，而且它完全运行在你的浏览器里。试试在线演示
Published:
Most enterprise security companies show up, ride one trend, and disappear in the next infrastructure cycle. Sean Xiang has been building Bloombase since January 2012, and the company has somehow been on the right side of every major infrastructure shift since, including the current AI accelerator era. That is not luck. That is a thesis.
Published:
大多数企业安全公司冒个泡、赶一波趋势，然后在下一个基础设施周期里消失。Sean Xiang 从 2012 年 1 月起就在打造 Bloombase，这家公司却阴差阳错站到了此后每一波重大基础设施转变的正确一边，包括当下的 AI 加速器时代。那不是运气，那是一套论点。
Published:
A web-based platform for facilitating cross-cultural communication and understanding.
Published:
一个基于 Web 的平台，用于促进跨文化交流与理解。
Published:
A widely shared post describes a brainstorming meeting at a robotics startup in Boston. A slide said “No stupid questions.” A Chinese employee interpreted it as “Do not ask stupid questions,” while American colleagues explained it as “Ask freely; no question will be judged stupid.” A related post claims that “Say it again?” is neutral when someone is not heard, whereas “What did you say?” is hostile.
Published:
一张流传截图讲了这样一件事：一家波士顿机器人创业公司开头脑风暴会，幻灯片写着 “No stupid questions.” 一位中国员工把它理解为“不要问蠢问题”，美国同事却说它的意思是“什么都可以问，没有问题会被当成蠢问题”。另一组讨论又声称：没听清时说 “Say it again?” 很中性，而 “What did you say?” 很不友好。
Published:
The silence surrounding the Alamo chapel in San Antonio belies the brutal, thirteen-day siege that transformed this former Spanish mission into the ultimate symbol of Texan independence. Standing before its weathered facade today, one can almost hear the echoes of a conflict that remains one of the most legendary chapters in American history.
Published:
Every product has its own pros and cons. Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
每一家都有每一家的优点和缺点。Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
Secretly recording a customer interview can turn a useful research habit into a privacy problem in one click.
Published:
偷偷录一次客户访谈，可能把一个正常的产品研究流程变成隐私和合规事故。
Published:
You think you picked your major rationally, but the choice itself is a bet.
Published:
第十三天的重点是让仪表盘更好看，并且把每条结论都说得有凭有据。
Published:
Most crypto dashboards either drown you in numbers or hide the signal behind a paywall. Crypto Dashboard is my attempt at a clean, AI-augmented view of the market that runs entirely in your browser. Try Live Demo
Published:
大多数加密货币仪表盘要么把你淹没在数字里，要么把信号藏在付费墙后面。Crypto Dashboard 是我对一种简洁、由 AI 增强的市场视图的尝试，而且它完全运行在你的浏览器里。试试在线演示
Published:
A comprehensive tool built with React, Vite, and Supabase to help individuals plan and navigate their careers in the AI industry.
Published:
这是一个基于 React、Vite 和 Supabase 构建的综合工具，帮助个人规划并导航自己在 AI 行业中的职业发展路径。
Published:
Sunk cost should never be a reason for a decision, and that includes your choice of major.
Published:
If I reopened the 2026 folder in 2030, I would check one thing first: did the work completed with AI change a real problem and leave evidence that others could verify and build upon?
Published:
假设在 2030 年重新打开 2026 年的文件夹，我会先查一件事：今天借助 AI 完成的工作，是否改变了一个现实问题，并留下别人能够复验和继承的证据。
Published:
At the end of AI’s compute road lies the ultimate solution to energy.
Published:
引言：AI 的算力尽头，是能源的终极解法。
Published:
An AI writing assistant that polishes academic text using DeepSeek-R1 and detects AI-generated content.
Published:
学术写作的严谨性不仅在于数据，更在于表达的精准，而 1AI-polish 通过集成 DeepSeek-R1 的推理能力，为研究者提供了一套集文本润色与 AI 检测于一体的深度协作系统，旨在让复杂的科研思想以更专业、更透明的方式呈现。
Published:
At today’s Silicon Valley alumni “Ask a Question” AI meet-up, what I wanted to talk about wasn’t how cool Claude Code is, but how it actually re-slices real work.
Published:
今天在硅谷校友会的「师问」AI 交流活动上，我想讲的不是 Claude Code 多酷，而是它到底怎么把真实工作重新拆开。
Published:
Scientists sell themselves too.
Published:
科学家也是出来卖的。
Published:
For technical people, a strong pull request on a fast-growing open-source repo can become better inbound marketing than another generic post about AI.
Published:
对技术人来说，在一个高速增长的开源仓库里做出高质量 PR，往往比再发一篇泛泛而谈的 AI 观点帖更像真正有效的 inbound marketing。
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
Published:
Should Silicon Valley engineers use OpenClaw?
Published:
硅谷大厂的工程师究竟需不需要用 OpenClaw？
Published:
Every product has its own pros and cons. Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
每一家都有每一家的优点和缺点。Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
In an era where AI can write code in seconds, I just learned the hard way that blindly moving fast is actually slowing me down.
Published:
在这个闭着眼睛就能让 AI 写代码的时代，我今天结结实实地挨了一锤。
Published:
My hamster’s cage broke.
Published:
我的仓鼠笼子坏了。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
Most VCs and accelerators talk about supporting founders. Kehan Dong specifically supports the kind of founder who started building at 16, when nobody told them they were allowed to, and it turns out that population is dramatically underserved.
Published:
大多数 VC 和孵化器都谈支持创始人。Kehan Dong 专门支持那种 16 岁就开始动手建造、而根本没人告诉过他们可以这么做的创始人，事实证明这个群体被严重忽视了。
Published:
When you join a lab, you do not just get a research direction. You inherit a 30-year codebase that is currently running inside the U.S. National Water Model and was on the critical path forecasting Hurricane Harvey. That is what working with Zong-Liang Yang at the Jackson School of Geosciences actually looks like.
Published:
加入实验室以后，你会接过一个研究方向，也会继承一套已有 30 年历史的代码库。它目前运行在美国国家水模型中，也曾处在飓风 Harvey 预报工作的关键路径上。这就是在 Jackson School of Geosciences 与 Zong-Liang Yang 一起工作的真实样子。
Published:
NASA’s Future Investigators in NASA Earth and Space Science and Technology (FINESST) is one of the most underused fellowships among US graduate students. Most PhD students have never heard of it, and the ones who have often miss the deadline because the proposal expectations aren’t obvious from the call alone. This repo is a curated guide of links, tips, and examples to help you write a competitive FINESST proposal.
Published:
NASA 的未来地球与空间科技研究者（FINESST）项目，是美国研究生中最被低估的奖学金之一。这个仓库是一份精心整理的链接、技巧与示例指南，帮助你写出一份有竞争力的 FINESST 提案。
Published:
Dr. Zong-Liang Yang’s Land Environment and Atmospheric Dynamics (LEAD) Group at UT-Austin employs satellite remote sensing, earth system modeling, and high-performance computing to advance understanding of Earth system sciences.
Published:
UT Austin 宗良杨博士（Dr. Zong-Liang Yang）领导的陆地环境与大气动力学研究组（LEAD Group）利用卫星遥感、地球系统建模和高性能计算，推动人类对地球系统科学的理解。
Published:
AlphaEarthHack is the project our team built for the UT Austin Geoscience Hackathon. The goal: see how far we could push AI on Earth system data inside a single weekend. Try Live Demo
Published:
AlphaEarthHack 是我们团队为德州大学奥斯汀分校地球科学黑客松打造的项目。目标只有一个：看看在一个周末内，我们能在地球系统数据上把 AI 推到多远。试试在线演示
Published:
Call it a (10^5) gap if you like, once you put semantic entropy and effective training samples on the board, the difference between internet-scale text and embodied robotics data stops sounding like rhetorical inflation and starts sounding almost conservative. You just have to separate raw physics bits from semantics a learner can actually use.
Published:
「十万倍」这个说法乍听像聊天里随手甩出去的量级，但如果把它放回语义信息熵和有效训练样本这两把尺子下面，它非但不夸张，反而可能是保守的。关键是你得把「原始物理比特」和「带语义的、可被学习的数据」分开看。
Published:
Decades of Earth-system modeling judgment live in PDFs, mailing lists, and senior researchers’ heads, and none of it is loadable by an AI coding agent. earth-space-ai.org is an attempt to fix that.
Published:
数十年来积累的地球系统建模判断散落在 PDF、邮件列表和资深研究者的头脑里，AI 编程智能体无法直接加载这些知识。earth-space-ai.org 试图解决这个问题。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
The roar of cooling systems in the Texas Advanced Computing Center (TACC) isn’t just noise—it’s the power of supercomputers translating the overwhelming dimensionality of climate data into something students can finally see and understand.
Published:
德州高级计算中心（TACC）里冷却系统的轰鸣不只是噪音——那是超级计算机的力量，正在把气候数据那令人不知所措的多维性，转化成学生最终能看见、能理解的东西。
Published:
The scary part is not that AI agents fail on Earth system model code, but that they can fail while looking almost right.
Published:
最可怕的部分不在于 AI 智能体在处理地球系统模型代码时失败了，而是它们可能在看似正确的代码中失败。
Published:
Most land surface models treat a tree like a passive straw. Water comes in at the roots, water leaves at the leaves, end of story. Ashley Matheny’s research basically says no, a tree is an active hydraulic system with storage, capacitance, and a strategy, and if you do not model it that way you are going to be wrong about drought.
Published:
大多数陆地表面模型都把一棵树当成一根被动的吸管。水从根进来，水从叶出去，故事就这么简单。Ashley Matheny 的研究基本上在说：不对，树是一个带有储水、电容和策略的活跃水力系统，如果你不这样建模，你在干旱问题上就会犯错误。
Published:
What Happens When You Just Give People Money? Sam Altman’s UBI Experiment
Published:
直接发钱会怎样？奥特曼的全民基本收入实验
Published:
The signal truly worth heeding is not that AI companies are growing too slowly, but that they are growing fast while cash-flow pressure keeps mounting.
Published:
真正值得警惕的信号不是 AI 公司增长太慢，而是它们增长得很快，现金流压力却仍在放大。
Published:
I joined Tsinghua University’s seminar series on the joint cultivation of top innovative talent as an invited student speaker. I shared a student’s view of cross-institution research training and the conditions that help young researchers pursue difficult questions.
Published:
我受邀以学生代表身份参加清华大学拔尖创新人才联合培养系列研讨会。发言从学生视角出发，谈跨校科研培养，也谈年轻研究者追问困难问题时需要怎样的环境。
Published:
Tsinghua News republished the People’s Daily Overseas Edition feature on Shenzhen X-Institute. The story describes the institute’s project-based education model and includes my experience of exploring research there as a USTC undergraduate.
Published:
清华新闻网转载了《人民日报海外版》关于深圳零一学院的报道。文章介绍了零一学院的项目制培养方式，也写到我在中国科大读本科时进入零一学院、尝试不同研究方向的经历。
Published:
The People’s Daily Overseas Edition featured Shenzhen X-Institute and its project-based approach to developing young innovators. I spoke about arriving at the institute as a USTC undergraduate, trying research across disciplines, and finding a community where students actively pursued their own questions.
Published:
《人民日报海外版》报道了深圳零一学院以项目制培养创新人才的探索。我在报道中回顾了自己从中国科大进入零一学院后，跨学科尝试科研、逐渐找到方向的经历，也谈到身边同学主动做项目、追问题的氛围。
Published:
Linchuan No. 1 Middle School published a congratulatory note when I was admitted to the University of Science and Technology of China. It records the transition from high school to the next stage of my education.
Published:
临川一中在我被中国科学技术大学录取后发布了祝贺文章。这篇短文记录了我从高中走向下一阶段学习的节点。
Published:
The article 教育变革的技术奇点时刻到来了吗？ discusses whether technology is changing the conditions of education. My contribution draws on my experience as a USTC student participating in Shenzhen X-Institute’s project-based research environment.
Published:
《教育变革的技术奇点时刻到来了吗？》讨论技术是否正在改变教育的基本条件。我结合自己在中国科大读书、参加深圳零一学院项目制研究的经历，提供了一名学生的视角。
Published:
China Science Daily’s WeChat edition tells how students who felt out of place in conventional university life found peers at Shenzhen X-Institute. My part follows the move from a difficult first year to leading a 3D-chip cooling project and completing the Senior Undergraduate Researcher Program.
Published:
《中国科学报》的微信版文章记录了一群不太适应传统大学生活的学生，如何在深圳零一学院遇见同路人。我的部分从大学第一年的低谷写起，讲到带队研究三维芯片散热，并完成高年级本科生研究计划的经历。
Published:
China Science Daily profiled Shenzhen X-Institute’s approach to innovation education. The article follows my path from a difficult first year at university to a student-led 3D-chip cooling project, then situates that experience in an education model built around questions, mentors, peers, and room to experiment.
Published:
《中国科学报》报道了深圳零一学院对创新人才培养的探索。文章写到我在大学第一年经历低谷，后来参加零一学院暑校，与队友从问题出发完成三维芯片散热项目，逐渐找到科研兴趣的过程。
Published:
The sense of agency that took me more than twenty years to grow, these seventeen and eighteen year old students got far more easily, just by being at Stanford, being in America, or maybe going to high school in Shanghai first.
Published:
I realized I had a serious misunderstanding in my “classes are useless” view.
Published:
Rereading my old journals today, from nuclear fusion in physics to the games humans play, I keep finding the same thing. Whether it is a startup or a whole industry, what matters is a force that rises on its own and then keeps itself going.
Published:
After high school, I think sitting in class was useless almost everywhere, in the mathematician’s sense of “a.e.”
Published:
School rankings are a single number standing in for a multi-dimensional decision. School Evaluator is a small web tool that lets you score a school across the dimensions that actually shape your life there. Try Live Demo
Published:
学校排名只是一个用单一数字去代表多维决策的替代品。School Evaluator 是一款小巧的网页工具，让你能在一个真正决定你在那里生活的维度上为学校打分。试试在线演示
Published:
In April 2024 I was the only undergraduate invited to speak among more than 40 scholars.
Published:
I was just thinking about why I can be polite to my brokerage’s customer service, but should not be polite to a middleman.
Published:
Earlier today I published a piece called “Four Ego Mistakes I Made as a 22-Year-Old Founder.” Then I ran it through Gemini 2.5 Pro as an independent reviewer. Gemini’s verdict was that the essay is itself an ego move. I think Gemini is mostly right.
Published:
今天早些时候我发了一篇叫《22 岁 founder 的四个 ego 错误》。然后我把这篇喂给了 Gemini 2.5 Pro 当独立审稿人。Gemini 的判决是：这篇文章本身就是一个 ego 动作。我觉得 Gemini 大体是对的。
Published:
When you send an urgent email asking for help, “quick question” gets a reply more easily than “emergency.”
Published:
This is the worst startup product I have seen.
Published:
Call it a (10^5) gap if you like, once you put semantic entropy and effective training samples on the board, the difference between internet-scale text and embodied robotics data stops sounding like rhetorical inflation and starts sounding almost conservative. You just have to separate raw physics bits from semantics a learner can actually use.
Published:
Why do some people who seem to work less than you succeed earlier? Because success is a question of timing first, and ability second.
Published:
When you send an urgent email asking for help, “quick question” gets a reply more easily than “emergency.”
Published:
At the end of AI’s compute road lies the ultimate solution to energy.
Published:
引言：AI 的算力尽头，是能源的终极解法。
Published:
In financial markets, we are forever hunting for the kind of “choke point” that everyone calls a stranglehold.
Published:
在金融市场里，我们永远在寻找那种被称为「卡脖子」的咽喉要道。
Published:
The most important measure of work may be the energy it consumes, more than the time it takes.
Published:
衡量工作效率，最重要的指标或许不是投入的时间，而是消耗的能源。
Published:
When something urgent lands at work and there is no time to prepare, so you stay up all night, that is a black swan.
Published:
Prompt Caching Is Not Technical Debt. It’s what the math demands.
Published:
很多人把 prompt caching 看成一个省钱 hack，但我觉得这个判断刚好反了。
Published:
Most enterprise security companies show up, ride one trend, and disappear in the next infrastructure cycle. Sean Xiang has been building Bloombase since January 2012, and the company has somehow been on the right side of every major infrastructure shift since, including the current AI accelerator era. That is not luck. That is a thesis.
Published:
大多数企业安全公司冒个泡、赶一波趋势，然后在下一个基础设施周期里消失。Sean Xiang 从 2012 年 1 月起就在打造 Bloombase，这家公司却阴差阳错站到了此后每一波重大基础设施转变的正确一边，包括当下的 AI 加速器时代。那不是运气，那是一套论点。
Published:
A salary can arrive on time every two weeks while your ability to leave the job gets smaller every year.
Published:
工资每两周准时到账，一个人离开这份工作的能力，却可能一年比一年小。
Published:
The opportunity is not the news itself. The opportunity is the friction created by the news.
Published:
Most VCs and accelerators talk about supporting founders. Kehan Dong specifically supports the kind of founder who started building at 16, when nobody told them they were allowed to, and it turns out that population is dramatically underserved.
Published:
大多数 VC 和孵化器都谈支持创始人。Kehan Dong 专门支持那种 16 岁就开始动手建造、而根本没人告诉过他们可以这么做的创始人，事实证明这个群体被严重忽视了。
Published:
Research, stock trading, and startups all share a common pattern: success comes from combining a defensible core competency with the ability to ride trending waves. You don’t need exhaustive preparation anymore. You need methodology and the ability to produce content when it matters. When the right moment arrives, you strike.
Published:
The more I think about it, the more I believe that doing business, commerce and trade, is the most fundamental form of charity.
Published:
The scary part is not that AI agents fail on Earth system model code, but that they can fail while looking almost right.
Published:
最可怕的部分不在于 AI 智能体在处理地球系统模型代码时失败了，而是它们可能在看似正确的代码中失败。
Published:
This afternoon I paid 80 dollars at an animal shelter in Austin and brought home a three-year-old orange cat.
Published:
Two years later, I am back in this familiar city.
Published:
The sense of agency that took me more than twenty years to grow, these seventeen and eighteen year old students got far more easily, just by being at Stanford, being in America, or maybe going to high school in Shanghai first.
Published:
As we walked out of the library, an elderly white woman with silver hair saw us and said “excuse me” three times. She spoke to us with real pride, took out her phone, and said she had seen the two of us taking photos with the frog statue out front. Every Christmas, she told us, that frog gets dressed up in a Christmas outfit, look, just like this! Then she asked, very sincerely, whether we would still be here at Christmas, and how we had even heard of Atherton, because the place is so unremarkable, so quiet, so low-key. We told her we would probably each be heading home by then. She added that her son had been learning Chinese since he was a little kid, and that he speaks it really well.
Published:
On the way home from SFO, I was actually pretty sad.
Published:
Walking University Avenue in Palo Alto today, I finally saw what many Chinese students get wrong, we treat Silicon Valley as a destination.
Published:
When you send an urgent email asking for help, “quick question” gets a reply more easily than “emergency.”
Published:
A few days ago Yann LeCun announced his new role in a LinkedIn post, something almost unthinkable in China.
Published:
There is no top-tier, universally applicable methodology, except one. Go deep in one industry, build personal influence, build a personal brand.
Published:
A year and a half into my PhD in the US, the best training I have received is learning to tell a good story.
Published:
I realized I had a serious misunderstanding in my “classes are useless” view.
Published:
I have a theory that the way our bodies are built is adapted to the ancient world.
Published:
When something urgent lands at work and there is no time to prepare, so you stay up all night, that is a black swan.
Published:
If you want to feel your own growth directly, write a diary every day.
Published:
After high school, I think sitting in class was useless almost everywhere, in the mathematician’s sense of “a.e.”
Published:
A bit of history from a casual chat: in the US, science students used to get bigger stipends than business, law or medicine students, because science students were poor.
Published:
The more I think about it, the more I believe that doing business, commerce and trade, is the most fundamental form of charity.
Published:
Don’t expect your first video to go viral, and don’t expect your first sales trip to land ten million dollars, but you can ask what happens after the hundredth try.
Published:
Why do some people who seem to work less than you succeed earlier? Because success is a question of timing first, and ability second.
Published:
Another day of heavy internet surfing, and the biggest lesson was this: there is no investment strategy that lets you lie down and get financially free.
Published:
Why not just go for 20 million, or 30 million?
Published:
One cup of milk tea, a tip of zero selected on the screen, and the shop still added a mandatory 18%.
Published:
I was just thinking about why I can be polite to my brokerage’s customer service, but should not be polite to a middleman.
Published:
If all your practice comes from school exams, you will never get that many chances to practice.
Published:
Microsoft and OpenAI share the same end goal, freeing people, and the only real difference between them is how deep their lever reaches.
Published:
I have come to despise the word “learning.”
Published:
Most crypto dashboards either drown you in numbers or hide the signal behind a paywall. Crypto Dashboard is my attempt at a clean, AI-augmented view of the market that runs entirely in your browser. Try Live Demo
Published:
大多数加密货币仪表盘要么把你淹没在数字里，要么把信号藏在付费墙后面。Crypto Dashboard 是我对一种简洁、由 AI 增强的市场视图的尝试，而且它完全运行在你的浏览器里。试试在线演示
Published:
Your next cloud instance is alive, it needs to be fed every three days, and it will be dead in six months whether you cancel or not.
Published:
第十三天的重点是让仪表盘更好看，并且把每条结论都说得有凭有据。
Published:
I have come to despise the word “learning.”
Published:
On the way home from SFO, I was actually pretty sad.
Published:
THIS IS A FAKE BLOG. The content below is fabricated and should not be cited or treated as a real interview or factual record.
Published:
胡晨曦的研究揭示，城市化不只是被动地应对极端天气，它还会主动重塑热带气旋如何向沿海特大城市倾泻暴雨。
Published:
I spent a whole evening with Gemini taking apart the Buy Borrow Die playbook that American billionaires run, expecting that with a little scaling down I could just copy the moves, and what I found instead was that almost every single move has an F-1 trapdoor underneath it, and the list of things I can actually do fits on one page.
Published:
我花了一个晚上跟 Gemini 把美国富豪那套 Buy Borrow Die 的玩法从头扒到尾，本来以为只要照抄就能复刻，结果一层层剥下来才发现，F-1 学生在这个剧本里几乎处处是坑，能用的招其实只有那么三两个。
Published:
THIS IS A FAKE BLOG. The content below is fabricated and should not be cited or treated as a real interview or factual record.
Published:
胡晨曦的研究揭示，城市化不只是被动地应对极端天气，它还会主动重塑热带气旋如何向沿海特大城市倾泻暴雨。
Published:
On the way home from SFO, I was actually pretty sad.
Published:
An AI writing assistant that polishes academic text using DeepSeek-R1 and detects AI-generated content.
Published:
学术写作的严谨性不仅在于数据，更在于表达的精准，而 1AI-polish 通过集成 DeepSeek-R1 的推理能力，为研究者提供了一套集文本润色与 AI 检测于一体的深度协作系统，旨在让复杂的科研思想以更专业、更透明的方式呈现。
Published:
A web-based platform for facilitating cross-cultural communication and understanding.
Published:
一个基于 Web 的平台，用于促进跨文化交流与理解。
Published:
你好，我是 Koutian。第十四天，我们把雷达的几个老毛病修好了，也让它更像一个会自己干活的工具。
Published:
The roar of cooling systems in the Texas Advanced Computing Center (TACC) isn’t just noise—it’s the power of supercomputers translating the overwhelming dimensionality of climate data into something students can finally see and understand.
Published:
德州高级计算中心（TACC）里冷却系统的轰鸣不只是噪音——那是超级计算机的力量，正在把气候数据那令人不知所措的多维性，转化成学生最终能看见、能理解的东西。
Published:
I plugged a Western Digital external drive full of data into a MacBook Air, tried to copy a file onto it, and nothing happened. No error dialog, no progress bar, just a drive that would let me read everything and write nothing. My first instinct was that something was broken, or that I needed to fix permissions. Both were wrong, and chasing the wrong explanation almost led me to permanently downgrade the security of the whole laptop.
Published:
我把一块装满数据的西部数据（WD）移动硬盘插到 MacBook Air 上，想往里拷一个文件，结果什么都没发生。没有报错弹窗，没有进度条，就是一块能读出所有东西、却一个字节都写不进去的硬盘。我的第一反应是它坏了，或者是我得去修一下权限。这两个判断都是错的，而且顺着错误的解释找下去，差点让我把整台笔记本的安全性永久降级。
| Warren Buffett: The “Open Source” Leader Half a Century Before GitHub | 巴菲特：比GitHub早了半个世纪的“开源”运动领袖 |
Published:
In an era driven by code, collaboration, and transparency, GitHub has become synonymous with the “open source” spirit. But if we turn our gaze to the world of finance, we find a pioneer of “open source” who predates the birth of GitHub by a full half-century. His name is Warren Buffett.
Published:
在当今这个由代码、协作和透明度驱动的时代，GitHub 成为了”开源”精神的代名词。但如果我们将目光投向金融界，会发现一位”开源”的先行者，他比 GitHub 的诞生早了整整半个世纪。他就是沃伦·巴菲特。
Published:
People love saying that ten years to a billion-dollar valuation is the normal pace for a unicorn, neither fast nor slow. But once you actually crunch the compound numbers, you realize what hides behind that word “normal” is a beast that doubles in value every single year.
Published:
The US is full of small traps that nobody warns you about until after you’ve fallen into one. US Survival Guide is a handbook for Chinese students and newcomers on identifying risk early, knowing your rights, and building a financial defense before you need it. Read the Guide
Published:
美国布满了各式各样的暗坑，在你掉进去之前，没人会提醒你。US Survival Guide 是一本写给中国学生和新移民的手册，教你及早识别风险、了解自己的权利，并在你需要之前先筑好财务防线。阅读指南
Published:
NASA’s Future Investigators in NASA Earth and Space Science and Technology (FINESST) is one of the most underused fellowships among US graduate students. Most PhD students have never heard of it, and the ones who have often miss the deadline because the proposal expectations aren’t obvious from the call alone. This repo is a curated guide of links, tips, and examples to help you write a competitive FINESST proposal.
Published:
NASA 的未来地球与空间科技研究者（FINESST）项目，是美国研究生中最被低估的奖学金之一。这个仓库是一份精心整理的链接、技巧与示例指南，帮助你写出一份有竞争力的 FINESST 提案。
Published:
Call it a (10^5) gap if you like, once you put semantic entropy and effective training samples on the board, the difference between internet-scale text and embodied robotics data stops sounding like rhetorical inflation and starts sounding almost conservative. You just have to separate raw physics bits from semantics a learner can actually use.
Published:
You can study fluid mechanics in five different countries before you turn 30, work on petroleum reservoirs and tectonophysics and planetary ice on the same week, and somehow end up at a Centennial Chair in Geophysics. Marc Hesse did exactly that, and the through-line is more interesting than any of the individual stops.
Published:
你可以在 30 岁前在五个不同国家学流体力学，同一周里既研究油气储层又研究构造物理和行星冰，最后竟然坐上地球物理学百年讲席。Marc Hesse 就是这么做的，而他背后那条主线比任何一个单独的站点都更有意思。
Published:
Rereading my old journals today, from nuclear fusion in physics to the games humans play, I keep finding the same thing. Whether it is a startup or a whole industry, what matters is a force that rises on its own and then keeps itself going.
Published:
Soy milk is coffee. Youtiao is a croissant.
Published:
咖啡也是豆浆，牛奶也是豆浆。
Published:
In October 2023 I asked ChatGPT an over-engineered question: how do you fry rice so that every grain of rice ends up bonded to egg, with no bare rice grains and no isolated clumps of egg sitting off on their own? I saved that conversation as an HTML file, dropped it in my Downloads folder, and did not look at it again for almost three years. Last week I finally turned it into an actual open-source repo, and going back through the original conversation to build it was more interesting than I expected.
Published:
2023 年 10 月，我问了 ChatGPT 一个过度工程化的问题：怎么炒蛋炒饭，才能让每一粒米都裹上蛋，既没有裸露的白米粒，也没有单独抱团、没沾到米的蛋碎？那次对话我存成了一个 HTML 文件，扔进 Downloads 文件夹，然后差不多三年没再打开过。上周我终于把它整理成了一个真正的开源仓库，回头重读那段 2023 年的对话、把它做成代码的过程，比我预想的有意思得多。
Published:
I’ve been playing with mixed-language builds (Fortran calling into C++ and C via iso_c_binding) and wanted a demo that was more interesting than “add two numbers across languages.” So I built fortran-zip-bomb: a small program that generates a genuine ZIP bomb — a small archive that expands into a much larger file on decompression.
Published:
我一直在玩混合语言构建（Fortran 通过 iso_c_binding 调用 C++ 和 C），想要一个比”跨语言把两个数加起来”更有意思的演示。于是我建了 fortran-zip-bomb：一个小程序，生成一个真正的 ZIP 炸弹——一个解压时会扩展成大得多的文件的小压缩包。
Published:
The same way text foundation models ate NLP and image models ate computer vision, someone is going to build the foundation model that eats spatial reasoning. Gengchen Mai is one of the people taking that bet seriously, and his SEAI Lab at UT Austin is one of the places it is being built.
Published:
文本基础模型已经改变了 NLP，图像模型也改变了计算机视觉。空间推理领域同样会出现自己的基础模型。Gengchen Mai 正在研究这一方向，他在 UT Austin 的 SEAI Lab 也开展相关工作。
Published:
Most enterprise security companies show up, ride one trend, and disappear in the next infrastructure cycle. Sean Xiang has been building Bloombase since January 2012, and the company has somehow been on the right side of every major infrastructure shift since, including the current AI accelerator era. That is not luck. That is a thesis.
Published:
大多数企业安全公司冒个泡、赶一波趋势，然后在下一个基础设施周期里消失。Sean Xiang 从 2012 年 1 月起就在打造 Bloombase，这家公司却阴差阳错站到了此后每一波重大基础设施转变的正确一边，包括当下的 AI 加速器时代。那不是运气，那是一套论点。
Published:
Earlier today I published a piece called “Four Ego Mistakes I Made as a 22-Year-Old Founder.” Then I ran it through Gemini 2.5 Pro as an independent reviewer. Gemini’s verdict was that the essay is itself an ego move. I think Gemini is mostly right.
Published:
今天早些时候我发了一篇叫《22 岁 founder 的四个 ego 错误》。然后我把这篇喂给了 Gemini 2.5 Pro 当独立审稿人。Gemini 的判决是：这篇文章本身就是一个 ego 动作。我觉得 Gemini 大体是对的。
Published:
Research date: 2026-05-26
Published:
研究时间：2026-05-26
Published:
When Large Language Models (LLMs) hallucinate, we often treat it as an engineering bug to be fixed. But what if hallucinations are not a flaw, but a mathematical inevitability? A recent interdisciplinary discussion revealed a profound connection between the Gibbs phenomenon in Fourier analysis and the fundamental limitations of neural networks.
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
Published:
I want the people close to me to live well. It is easy to let that wish turn into deciding what living well should mean for them.
Published:
我希望亲近的人过得好，也很容易顺着这个愿望，替他们想好怎样才算过得好。
Published:
Securing up to $150,000 in research funding over three years can define a graduate career, and NASA’s FINESST program is the primary vehicle for that transformation. This guide distills the complex application process into actionable strategies for Earth and Space Science researchers seeking to join the next generation of Future Investigators.
Published:
每年 5 万美元且连续 3 年的科研资助，让 NASA FINESST 项目成为地球与空间科学博士生必须争取的黄金机会。这份指南旨在将复杂的申请流程拆解为可操作的策略，帮助下一代“未来研究员”（Future Investigators）在激烈的竞争中脱颖而出。
Published:
If I reopened the 2026 folder in 2030, I would check one thing first: did the work completed with AI change a real problem and leave evidence that others could verify and build upon?
Published:
假设在 2030 年重新打开 2026 年的文件夹，我会先查一件事：今天借助 AI 完成的工作，是否改变了一个现实问题，并留下别人能够复验和继承的证据。
Published:
In the future world, everyone will have their own AI agent assistant.
Published:
在未来的世界，每个人都会有自己的 AI agent 助理。
Published:
Sometimes a piece of old science fiction stays in your head not because of the gadgets, but because it got the emotional structure of the future right.
Published:
有些旧科幻会一直留在脑子里，不只是因为它写了什么炫酷科技，而是因为它提前抓住了未来的情感结构。
Published:
What Happens When You Just Give People Money? Sam Altman’s UBI Experiment
Published:
Every knowledge worker is now competing with an LLM for some part of their job. AI Survival Guide is a handbook for figuring out which parts, what to do about it, and how to come out the other side better at your craft instead of replaced by it.
Published:
如今，每一位知识工作者都在用某项工作的某些部分与 LLM 竞争。AI Survival Guide 是一本手册，帮你弄清是哪些部分、该怎么应对，以及如何带着更强的技艺从这场竞争的另一头走出来，而不是被它取代。
Published:
What do a 7,169-yuan Chinese animated film and a 483-dollar GameStop short squeeze have in common? More than you would think.
Published:
一部票房 7169 元的国产动画，和一只冲到 483 美元的 GameStop 股票，有什么共同点？比你想象的要多。
Published:
Virtual cell models powered by neuromorphic computing and gene language models represent one of the most capital-intensive and scientifically ambitious convergences in biotech AI. If you’re evaluating this space as a VC intern, you need to understand three core components: what these systems actually do, why the market is moving now, and where the investable opportunities lie.
Published:
I am deeply honored to join my colleagues in contributing to the newly published paper, “On the Ethics of Generative GeoAI: Explainability, Bias, Hallucination, Accountability, Privacy, and Trust,” specifically focusing on the section regarding GeoAI Trust (Section 8).
Published:
我非常荣幸能够参与到这篇新发表的论文 《On the Ethics of Generative GeoAI: Explainability, Bias, Hallucination, Accountability, Privacy, and Trust》 的合作中，并作为共同作者之一，贡献了关于 GeoAI信任（GeoAI Trust） 的章节（第八章）。
Published:
The same way text foundation models ate NLP and image models ate computer vision, someone is going to build the foundation model that eats spatial reasoning. Gengchen Mai is one of the people taking that bet seriously, and his SEAI Lab at UT Austin is one of the places it is being built.
Published:
文本基础模型已经改变了 NLP，图像模型也改变了计算机视觉。空间推理领域同样会出现自己的基础模型。Gengchen Mai 正在研究这一方向，他在 UT Austin 的 SEAI Lab 也开展相关工作。
Published:
AlphaEarthHack is the project our team built for the UT Austin Geoscience Hackathon. The goal: see how far we could push AI on Earth system data inside a single weekend. Try Live Demo
Published:
AlphaEarthHack 是我们团队为德州大学奥斯汀分校地球科学黑客松打造的项目。目标只有一个：看看在一个周末内，我们能在地球系统数据上把 AI 推到多远。试试在线演示
Published:
You can study fluid mechanics in five different countries before you turn 30, work on petroleum reservoirs and tectonophysics and planetary ice on the same week, and somehow end up at a Centennial Chair in Geophysics. Marc Hesse did exactly that, and the through-line is more interesting than any of the individual stops.
Published:
你可以在 30 岁前在五个不同国家学流体力学，同一周里既研究油气储层又研究构造物理和行星冰，最后竟然坐上地球物理学百年讲席。Marc Hesse 就是这么做的，而他背后那条主线比任何一个单独的站点都更有意思。
Published:
In my first post on this I described sending a patch through GitGitGadget and watching it land on the mailing list as v1. That post ended with the patch waiting for review. It has since merged. Now that I have seen the full lifecycle, from typo to master, here is what actually mattered.
Published:
在第一篇文章里，我写了怎么通过 GitGitGadget 把补丁发出去，它作为 v1 落在了邮件列表上。那篇文章结束在等待 review 的状态。现在它已经合并了。看完从一个拼写错误到进入 master 的完整过程，下面是我觉得真正重要的几点。
Published:
The patch I sent to the GitGitGadget doorstep ten days ago has been merged into master.
Published:
十天前我送到 GitGitGadget 门口的那个补丁，现在已经合并进 master 了。
Published:
Today I got my first Git patch onto the mailing list.
Published:
今天我差一点，就把自己的第一个 Git 项目补丁送进邮件列表了。
Published:
In my first post on this I described sending a patch through GitGitGadget and watching it land on the mailing list as v1. That post ended with the patch waiting for review. It has since merged. Now that I have seen the full lifecycle, from typo to master, here is what actually mattered.
Published:
在第一篇文章里，我写了怎么通过 GitGitGadget 把补丁发出去，它作为 v1 落在了邮件列表上。那篇文章结束在等待 review 的状态。现在它已经合并了。看完从一个拼写错误到进入 master 的完整过程，下面是我觉得真正重要的几点。
Published:
The patch I sent to the GitGitGadget doorstep ten days ago has been merged into master.
Published:
十天前我送到 GitGitGadget 门口的那个补丁，现在已经合并进 master 了。
Published:
Today I got my first Git patch onto the mailing list.
Published:
今天我差一点，就把自己的第一个 Git 项目补丁送进邮件列表了。
Published:
The opportunity is not the news itself. The opportunity is the friction created by the news.
Published:
新闻本身不是机会。新闻制造的摩擦才是机会。
Published:
For technical people, a strong pull request on a fast-growing open-source repo can become better inbound marketing than another generic post about AI.
Published:
对技术人来说，在一个高速增长的开源仓库里做出高质量 PR，往往比再发一篇泛泛而谈的 AI 观点帖更像真正有效的 inbound marketing。
| Warren Buffett: The “Open Source” Leader Half a Century Before GitHub | 巴菲特：比GitHub早了半个世纪的“开源”运动领袖 |
Published:
In an era driven by code, collaboration, and transparency, GitHub has become synonymous with the “open source” spirit. But if we turn our gaze to the world of finance, we find a pioneer of “open source” who predates the birth of GitHub by a full half-century. His name is Warren Buffett.
Published:
在当今这个由代码、协作和透明度驱动的时代，GitHub 成为了”开源”精神的代名词。但如果我们将目光投向金融界，会发现一位”开源”的先行者，他比 GitHub 的诞生早了整整半个世纪。他就是沃伦·巴菲特。
Published:
你好，我是 Koutian。第十五天，雷达第一次学会主动对外说话，不再只是等你来翻。
Published:
雷达现在能订阅了。第十二天，我们加了一个每天更新的 RSS 订阅源。
Published:
Why not just go for 20 million, or 30 million?
Published:
When gold gets too expensive relative to silver, you sell your gold and buy silver. When silver catches up, you swap back. This sounds like folk wisdom, but it’s one of the most time-tested strategies in precious metals.
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
Published:
NASA’s Future Investigators in NASA Earth and Space Science and Technology (FINESST) is one of the most underused fellowships among US graduate students. Most PhD students have never heard of it, and the ones who have often miss the deadline because the proposal expectations aren’t obvious from the call alone. This repo is a curated guide of links, tips, and examples to help you write a competitive FINESST proposal.
Published:
NASA 的未来地球与空间科技研究者（FINESST）项目，是美国研究生中最被低估的奖学金之一。这个仓库是一份精心整理的链接、技巧与示例指南，帮助你写出一份有竞争力的 FINESST 提案。
Published:
Securing up to $150,000 in research funding over three years can define a graduate career, and NASA’s FINESST program is the primary vehicle for that transformation. This guide distills the complex application process into actionable strategies for Earth and Space Science researchers seeking to join the next generation of Future Investigators.
Published:
每年 5 万美元且连续 3 年的科研资助，让 NASA FINESST 项目成为地球与空间科学博士生必须争取的黄金机会。这份指南旨在将复杂的申请流程拆解为可操作的策略，帮助下一代“未来研究员”（Future Investigators）在激烈的竞争中脱颖而出。
Published:
A PhD is hard. A PhD in 2026 — with the field moving faster than your committee can read — is a different kind of hard. AI PhD Survival Guide is a handbook for surviving and finishing an AI or ML PhD without burning out and without falling behind.
Published:
读博很难。2026 年读博——当领域进展快到你的委员会都读不过来时——是另一种难。AI PhD Survival Guide 是一本手册，教你如何在 AI 或机器学习博士研究中活下来并毕业，既不会燃烧殆尽，也不至于掉队。
Published:
School rankings are a single number standing in for a multi-dimensional decision. School Evaluator is a small web tool that lets you score a school across the dimensions that actually shape your life there. Try Live Demo
Published:
学校排名只是一个用单一数字去代表多维决策的替代品。School Evaluator 是一款小巧的网页工具，让你能在一个真正决定你在那里生活的维度上为学校打分。试试在线演示
Published:
Dr. Zong-Liang Yang’s Land Environment and Atmospheric Dynamics (LEAD) Group at UT-Austin employs satellite remote sensing, earth system modeling, and high-performance computing to advance understanding of Earth system sciences.
Published:
UT Austin 宗良杨博士（Dr. Zong-Liang Yang）领导的陆地环境与大气动力学研究组（LEAD Group）利用卫星遥感、地球系统建模和高性能计算，推动人类对地球系统科学的理解。
Published:
Hah, I can’t help but laugh. It just hit me: the first time I ever used Markdown was back when I was building a poetry society. And that’s also when I first learned about Git. Looking back now, if you put all the founding members of that poetry society together, you’d almost have two unicorns.
Published:
哎，我他妈笑了。我忽然想起来，我最早用 Markdown，就是之前创建诗社的时候。知道 Git，也是在那个时候。现在回头看，整个诗社的元老凑在一起，真的快有两个独角兽了。
Published:
If you want to feel your own growth directly, write a diary every day.
Published:
AlphaEarthHack is the project our team built for the UT Austin Geoscience Hackathon. The goal: see how far we could push AI on Earth system data inside a single weekend. Try Live Demo
Published:
AlphaEarthHack 是我们团队为德州大学奥斯汀分校地球科学黑客松打造的项目。目标只有一个：看看在一个周末内，我们能在地球系统数据上把 AI 推到多远。试试在线演示
Published:
My hamster’s cage broke.
Published:
我的仓鼠笼子坏了。
Published:
I plugged a Western Digital external drive full of data into a MacBook Air, tried to copy a file onto it, and nothing happened. No error dialog, no progress bar, just a drive that would let me read everything and write nothing. My first instinct was that something was broken, or that I needed to fix permissions. Both were wrong, and chasing the wrong explanation almost led me to permanently downgrade the security of the whole laptop.
Published:
我把一块装满数据的西部数据（WD）移动硬盘插到 MacBook Air 上，想往里拷一个文件，结果什么都没发生。没有报错弹窗，没有进度条，就是一块能读出所有东西、却一个字节都写不进去的硬盘。我的第一反应是它坏了，或者是我得去修一下权限。这两个判断都是错的，而且顺着错误的解释找下去，差点让我把整台笔记本的安全性永久降级。
Published:
Seizing the gateway to the human mind in the AI era is, at bottom, a secret war over token throughput power.
Published:
抢占 AI 时代的人类心智入口，说到底是一场关于 token 吞吐权力的隐秘战争。
Published:
When an ordinary-looking “heart age” questionnaire lets you calculate that your heart is younger than your actual age, you may need to re-examine the lifestyle habits you’ve been ignoring.
Published:
当一份看似普通的“心脏年龄”问卷让你算出自己的心脏比实际年龄还年轻时，你可能需要重新审视一下那些被你忽视的生活习惯了。
Published:
I have a theory that the way our bodies are built is adapted to the ancient world.
Published:
An elegant couple mood tracking website with self-hosted backend and SQLite, displaying daily mood records in GitHub contribution graph style.
Published:
这是一个优雅的情侣心情记录网站，采用自托管后端和 SQLite，以 GitHub contribution graph 风格展示每日心情记录。
Published:
Decades of Earth-system modeling judgment live in PDFs, mailing lists, and senior researchers’ heads, and none of it is loadable by an AI coding agent. earth-space-ai.org is an attempt to fix that.
Published:
数十年来积累的地球系统建模判断散落在 PDF、邮件列表和资深研究者的头脑里，AI 编程智能体无法直接加载这些知识。earth-space-ai.org 试图解决这个问题。
Published:
A university’s favorite performance is to wait until a person no longer needs a medal, then ceremonially drape them in a gown embroidered with the word “Doctor.”
Published:
大学最擅长的表演，莫过于在一个人不再需要勋章时，隆重地为他披上一件绣着「博士」二字的学位袍。
Published:
Original Publication Date: March 15, 2026
Published:
Original Publication Date: March 15, 2026
Published:
A PhD system built on a 19th-century apprenticeship model is failing by nearly every empirical measure — ~40–50% attrition, depression rates six times the general population, and tenure-track placement below 15% in many fields — while venture capital has spent four decades perfecting bilateral contracts that manage exactly the risks PhD programs ignore: information asymmetry, moral hazard, hold-up, and misaligned incentives. The literature across economics, education policy, signaling theory, and AI research converges on a striking conclusion: the structural tools to fix the PhD already exist in VC contract design, but academia has never imported them. This research compendium maps the evidentiary landscape across six domains to ground the argument.
Published:
Whether a civilization has logic is not decided by whether it produced a few brilliant arguers. Three tests decide it. Did anyone study inference itself and write books about it. Did the subject enter schools, with teachers who taught it and students who learned it. Did some institution, an examination above all, force generation after generation to master it. Greece passes all three tests. India passes all three. Ancient China fails every one.
Published:
一个文明算不算有逻辑，不看它出过几个会推理的天才，看三样东西。有没有人把推理本身当成研究对象，写出专门的书。这门学问进没进学校，有没有老师教，学生学。有没有一套制度逼着一代代人学它，用它，考试就是最硬的制度。拿这三样去量希腊，量印度，都量得过。拿来量古代中国，三样全空。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
The silence surrounding the Alamo chapel in San Antonio belies the brutal, thirteen-day siege that transformed this former Spanish mission into the ultimate symbol of Texan independence. Standing before its weathered facade today, one can almost hear the echoes of a conflict that remains one of the most legendary chapters in American history.
Published:
On May 17, 2025, I had about nine hours alone in Hong Kong between flights and trains, and I planned them around Victoria Harbour and the city’s richest commercial districts.
Published:
2025年5月17日，我一人途经香港，可游览时间约九个小时，目标是维多利亚港和香港最繁华的商业地段。
Published:
A university’s favorite performance is to wait until a person no longer needs a medal, then ceremonially drape them in a gown embroidered with the word “Doctor.”
Published:
大学最擅长的表演，莫过于在一个人不再需要勋章时，隆重地为他披上一件绣着「博士」二字的学位袍。
Published:
A simple, clean HTML page displaying email verification success message with responsive design and professional styling.
Published:
这是一个简洁干净的 HTML 页面，用于展示邮箱验证成功提示，具备响应式布局和专业风格设计。
Published:
I submitted our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation to Hugging Face Daily Papers on September 14, 2026, and readers voted it to the top of that day’s list. As of September 29 it has 171 upvotes, the most of the 26 papers posted that day.
Published:
2026 年 9 月 14 日，我把论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 提交到 Hugging Face Daily Papers，读者把它投上了当天榜首。截至 9 月 29 日，它获得 171 个赞，在当天 26 篇论文中排第一。
Published:
I ended up in the wrong NEXUS line at Vancouver International Airport. By the rules, the mistake was mine. By design, it was an almost predictable wrong turn.
Published:
The last mile of an AI-generated travel plan may be one missing Retry button.
Published:
This afternoon I paid 80 dollars at an animal shelter in Austin and brought home a three-year-old orange cat.
Published:
@elonmusk, asking for a friend, has Tesla FSD been tested in weather like this?
Published:
What do a 7,169-yuan Chinese animated film and a 483-dollar GameStop short squeeze have in common? More than you would think.
Published:
一部票房 7169 元的国产动画，和一只冲到 483 美元的 GameStop 股票，有什么共同点？比你想象的要多。
Published:
How can a generative-AI student society do more than host talks, repost news, and chase the latest model?
Published:
一个生成式 AI 学生社团，怎样才能不止于办讲座、转发新闻和追逐下一款模型？
Published:
The sharpest minds keep moving into AI labs, and it isn’t just for the money.
Published:
这是《颠覆性创新者的思维模式》的五个月后续研究。
Published:
Research date: 2026-05-26
Published:
研究时间：2026-05-26
Published:
If you ask most people where trees in California get their water in a drought, they will say “the soil.” It turns out a huge fraction of it comes from cracks in the bedrock underneath the soil, and Daniella Rempe is the person who put numbers on it.
Published:
如果你问大多数人，加州在干旱时树从哪里取水，他们会说”土壤”。事实证明，很大一部分水其实来自土壤下方岩石裂缝里的基岩，而 Daniella Rempe 就是把数字放到这件事上的人。
Published:
Day thirty-one of Benchmark Radar. Every release card now carries a one-line summary instead of a bare version tag, filing an issue takes at most two required fields, and both READMEs got short enough to read. Scoreboard: 107 stars, 20 forks.
Published:
Benchmark Radar 的第三十一天。每张发布卡片现在都带一行摘要，而不是一个光秃秃的版本号；提交 issue 最多填两个必填项；两份 README 也精简到读得下去。记分牌：107 颗星，20 个 fork。
Published:
Day twenty-eight of Benchmark Radar. We removed a second copy of the data, made the page say one thing to crawlers, and gave the work a citable name.
Published:
别给同一份真相留两个副本，否则它们迟早对不上。第二十八天我们做了三件事：把可生成的数据移出版本库，让页面只对爬虫说一个标题，给作品一个可引用的名字。先说几个词：真相来源是唯一可信的那份源文件，其余都由它生成；h1 是页面的主标题，爬虫期望整页只有一个；i18n 是国际化，让界面按语言显示；引用是你写论文时粘贴的那条参考文献。
Published:
Day twenty-seven of Benchmark Radar. We made the Today list load a page at a time, cut a title in half, and taught 21 benchmarks to show their real names.
Published:
一天 136 条还要一次全画出来，只会让首屏变慢。第二十七天我们做了三件事：让今日列表一页页加载，把标题砍半，让 21 个基准显示真名。先说几个词：分页是一长串内容分多页看完；首包是页面为了快而先加载的小数据包；SEO 是让搜索引擎看懂并收录网站的做法。
Published:
Hi, Koutian here. Day twenty-one was a fix day that touched many parts at once.
Published:
OpenReview 完成了认证接入，宽屏布局得以扩展，手机端的显示问题也修复了。第二十一天是横跨多个模块的修复日。
Published:
Hi, Koutian here. Day twenty taught the radar to speak Chinese and to summarize itself.
Published:
雷达开始说中文了。第二十天交付了完整的 zh 国际化词典、洞察块和完整的社交清单。
Published:
I spent a whole evening with Gemini taking apart the Buy Borrow Die playbook that American billionaires run, expecting that with a little scaling down I could just copy the moves, and what I found instead was that almost every single move has an F-1 trapdoor underneath it, and the list of things I can actually do fits on one page.
Published:
我花了一个晚上跟 Gemini 把美国富豪那套 Buy Borrow Die 的玩法从头扒到尾，本来以为只要照抄就能复刻，结果一层层剥下来才发现，F-1 学生在这个剧本里几乎处处是坑，能用的招其实只有那么三两个。
Published:
Every product has its own pros and cons. Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
每一家都有每一家的优点和缺点。Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
A laser lightsail accelerating to one percent of light speed turns about one percent of the beam into motion. The other ninety-nine percent leaves as light and does no work.
Published:
激光光帆加速到光速的百分之一，光束里大约只有百分之一变成了运动。剩下的百分之九十九以光的形式离开，不做功。
Published:
Today, alumnus Pan Jianwei sent his spiritual genes (his name) beyond the Earth.
Published:
今天，校友潘建伟把自己的精神基因（名字）传播到了地球之外。
Published:
A long time ago, I came across a book called Chase, Chance, and Creativity: The Lucky Art of Novelty（《追逐、机遇和创造力：新奇的幸运艺术》）. I no longer remember everything in it, but its central idea stayed somewhere in the back of my mind: luck and creativity are not entirely random. Chance may arrive unexpectedly, but we can still choose how often we encounter it and whether we are ready to recognize it.
Published:
很久以前，我偶然看到一本书，叫 Chase, Chance, and Creativity: The Lucky Art of Novelty（《追逐、机遇和创造力：新奇的幸运艺术》）。我已经不记得书里的全部内容了，但它的核心想法一直留在我脑海中的某个角落：运气和创造力并不完全是随机的。机遇也许会意外到来，但我们仍然可以选择自己遇见它的频率，以及当它出现时，我们是否已经准备好认出它。
Published:
Going around GitHub opening issues and pull requests is sperm donation. The same mechanism, running as-is.
Published:
在 GitHub 上到处开 issue、提 pull request，就是赛博捐精。同一套机制，原样运行。
Published:
Let me offer a comparison that’s a bit inappropriate but does capture a real structure of open-source collaboration:
Published:
我提一个不太恰当、但确实抓住了开源协作某种结构的比喻：
Published:
The clearest advice I recently received from a friend who has visited Japan several times was simple: plan the trip first.
Published:
When you go looking for a seat in the great theater of the AI era, built from PowerPoints, papers, and anxiety, it helps to look at which play each person on stage is performing.
Published:
当你在这个由PPT、Paper和焦虑构成的AI时代大剧院里找座位时，不如先看看台上的人都在演哪一出戏。
Published:
I was just thinking about why I can be polite to my brokerage’s customer service, but should not be polite to a middleman.
Published:
Call it a (10^5) gap if you like, once you put semantic entropy and effective training samples on the board, the difference between internet-scale text and embodied robotics data stops sounding like rhetorical inflation and starts sounding almost conservative. You just have to separate raw physics bits from semantics a learner can actually use.
Published:
The promise of AI-automated science is intoxicating: imagine machines that can generate hypotheses, design experiments, and publish papers while we sleep.
Published:
AI 自动化科研的愿景令人陶醉：想象机器能在我们睡觉时生成假设、设计实验、发表论文。然而，尽管关于“AI 科学家”的新闻铺天盖地，我们正在撞上一堵根本性的墙。问题不在于算力或数据集规模，而是某种更深刻的东西，根植于科学发现的本质和信息论之中。
Published:
I just remembered something sudden on the train: why are almost all the links on the internet that extremely saturated shade of blue?
Published:
我刚刚在火车上突然想起一件事：为什么互联网上的链接，几乎清一色都是那种饱和度极高的蓝色？
Published:
Most enterprise security companies show up, ride one trend, and disappear in the next infrastructure cycle. Sean Xiang has been building Bloombase since January 2012, and the company has somehow been on the right side of every major infrastructure shift since, including the current AI accelerator era. That is not luck. That is a thesis.
Published:
大多数企业安全公司冒个泡、赶一波趋势，然后在下一个基础设施周期里消失。Sean Xiang 从 2012 年 1 月起就在打造 Bloombase，这家公司却阴差阳错站到了此后每一波重大基础设施转变的正确一边，包括当下的 AI 加速器时代。那不是运气，那是一套论点。
Published:
In financial markets, we are forever hunting for the kind of “choke point” that everyone calls a stranglehold.
Published:
在金融市场里，我们永远在寻找那种被称为「卡脖子」的咽喉要道。
Published:
Most VCs and accelerators talk about supporting founders. Kehan Dong specifically supports the kind of founder who started building at 16, when nobody told them they were allowed to, and it turns out that population is dramatically underserved.
Published:
大多数 VC 和孵化器都谈支持创始人。Kehan Dong 专门支持那种 16 岁就开始动手建造、而根本没人告诉过他们可以这么做的创始人，事实证明这个群体被严重忽视了。
| The Mindset of Disruptive Innovators | What Did the People Who Changed the World Actually See? | An Epistemic Crisis in an Age of Abundance |
Published:
What do the people who change the world see that ordinary people can’t?
Published:
改变世界的人到底看到了什么是普通人看不到的？
Published:
Have you ever run into this while using AI?
Published:
你在用 AI 的时候，有没有遇到过这种情况？
Published:
Geepity turned our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation into a full interactive reading edition, walking through all nine chapters with live charts and a working BM25F demo.
Published:
Geepity 把我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 做成完整互动解读版，九章走完，配实时图表与可运行的 BM25F 演示。
Published:
A handcrafted website universe where every webpage, balloon, click, and line of code says “I love you” - designed exclusively for Momo. Visit Live Site
Published:
这是一个手工打造的网站小宇宙：每一个网页、气球、点击和每一行代码都在说“我爱你”——只为 Momo 而设计。访问在线站点
Published:
Here is a view I hold now: whatever you do, you look a lot like a speculator in a market.
Published:
How photographing a black hole and training GPT-4 are fundamentally the same process: extracting coherent structure from impossibly sparse frequency-domain measurements.
Published:
I ended up in the wrong NEXUS line at Vancouver International Airport. By the rules, the mistake was mine. By design, it was an almost predictable wrong turn.
Published:
Beyond the wayfinding problem that led me into the wrong NEXUS line, my connection at Vancouver International Airport left me with several observations unrelated to signs. I noticed an older-looking mix of people, a departures area that offered almost nowhere to work, and a visible contrast between people in Vancouver and Texas. This post records those observations. The full account of the NEXUS incident is in How I “Reasonably” Ended Up in the Wrong NEXUS Line at YVR.
Published:
The last mile of an AI-generated travel plan may be one missing Retry button.
Published:
I spent a whole evening with Gemini taking apart the Buy Borrow Die playbook that American billionaires run, expecting that with a little scaling down I could just copy the moves, and what I found instead was that almost every single move has an F-1 trapdoor underneath it, and the list of things I can actually do fits on one page.
Published:
The US is full of small traps that nobody warns you about until after you’ve fallen into one. US Survival Guide is a handbook for Chinese students and newcomers on identifying risk early, knowing your rights, and building a financial defense before you need it. Read the Guide
Published:
美国布满了各式各样的暗坑，在你掉进去之前，没人会提醒你。US Survival Guide 是一本写给中国学生和新移民的手册，教你及早识别风险、了解自己的权利，并在你需要之前先筑好财务防线。阅读指南
Published:
Checking flights today, I suddenly noticed: AI is the airline code for Air India.
Published:
今天看航班时，我突然发现：AI 是 Air India（印度航空）的航空公司代码。
Published:
From the vast datasets of Earth System Models to the specialized niches of high-performance computing, this space documents my journey as a PhD student at UT Austin pushing the boundaries of Geological and Earth Sciences. This website is more than just a portfolio—it’s a hub where data-driven climate science meets the practical challenges of modern research.
Published:
在 UT Austin 攻读地球科学博士学位的过程中，我始终在试图寻找复杂气候模型与真实世界影响之间的联结。这个学术网站不仅是我研究、论文和项目经历的展示入口，更是我记录如何利用数据驱动的科学方法去理解地球系统演变的思考空间。
Published:
I spent a whole evening with Gemini taking apart the Buy Borrow Die playbook that American billionaires run, expecting that with a little scaling down I could just copy the moves, and what I found instead was that almost every single move has an F-1 trapdoor underneath it, and the list of things I can actually do fits on one page.
Published:
When gold gets too expensive relative to silver, you sell your gold and buy silver. When silver catches up, you swap back. This sounds like folk wisdom, but it’s one of the most time-tested strategies in precious metals.
Published:
Another day of heavy internet surfing, and the biggest lesson was this: there is no investment strategy that lets you lie down and get financially free.
Published:
In financial markets, we are forever hunting for the kind of “choke point” that everyone calls a stranglehold.
Published:
在金融市场里，我们永远在寻找那种被称为「卡脖子」的咽喉要道。
Published:
The clearest advice I recently received from a friend who has visited Japan several times was simple: plan the trip first.
Published:
“这b人生过的值不值” — roughly, “was this damn life worth it?” — is a phrase that does a lot of emotional work on the Chinese internet. ZuoGuoHuaDiao (做过划掉) is a small web tool that takes the phrase seriously and tries to answer it with numbers. Try Live Demo
Published:
“这b人生过的值不值”——简单说，就是”这该死的一生过得值不值？”——这句在中国互联网上承载了大量情感的话语。ZuoGuoHuaDiao（做过划掉）是一款小巧的网页工具，它认真对待这句话，并试图用数字来回答它。试试在线演示
Published:
Most people think of summer vacation as time off. I started wondering whether that framing undersells it. Summer Calculator is a small web tool that estimates the full value of your summer — money, learning, relationships, health — not just the days you didn’t work. Try Live Demo
Published:
大多数人把暑假当作休息时间。我开始怀疑这种框架是否低估了它。Summer Calculator 是一款小巧的网页工具，用来估算你暑假的完整价值——金钱、学习、人际关系、健康——而不仅仅是你没工作的那些日子。试试在线演示
Published:
A handcrafted website universe where every webpage, balloon, click, and line of code says “I love you” - designed exclusively for Momo. Visit Live Site
Published:
这是一个手工打造的网站小宇宙：每一个网页、气球、点击和每一行代码都在说“我爱你”——只为 Momo 而设计。访问在线站点
Published:
Lay out every hyped multiAgent product on a table in front of Steve Jobs and he would probably squint, frown, and ask one question, why on earth are you showing the user any of this.
Published:
把现在市面上所有被吹得天花乱坠的multiAgent产品摆在乔布斯面前，他大概率会皱着眉头反问一句，你为什么要让用户看到这些东西。
Published:
If you want to feel your own growth directly, write a diary every day.
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
AlphaEarthHack is the project our team built for the UT Austin Geoscience Hackathon. The goal: see how far we could push AI on Earth system data inside a single weekend. Try Live Demo
Published:
AlphaEarthHack 是我们团队为德州大学奥斯汀分校地球科学黑客松打造的项目。目标只有一个：看看在一个周末内，我们能在地球系统数据上把 AI 推到多远。试试在线演示
Published:
Every product has its own pros and cons. Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
每一家都有每一家的优点和缺点。Cursor: “extraordinarily productive”; Kiro: “spec-driven”; Antigravity: “agent-first”.
Published:
The same way text foundation models ate NLP and image models ate computer vision, someone is going to build the foundation model that eats spatial reasoning. Gengchen Mai is one of the people taking that bet seriously, and his SEAI Lab at UT Austin is one of the places it is being built.
Published:
文本基础模型已经改变了 NLP，图像模型也改变了计算机视觉。空间推理领域同样会出现自己的基础模型。Gengchen Mai 正在研究这一方向，他在 UT Austin 的 SEAI Lab 也开展相关工作。
Published:
Lay out every hyped multiAgent product on a table in front of Steve Jobs and he would probably squint, frown, and ask one question, why on earth are you showing the user any of this.
Published:
把现在市面上所有被吹得天花乱坠的multiAgent产品摆在乔布斯面前，他大概率会皱着眉头反问一句，你为什么要让用户看到这些东西。
Published:
Every knowledge worker is now competing with an LLM for some part of their job. AI Survival Guide is a handbook for figuring out which parts, what to do about it, and how to come out the other side better at your craft instead of replaced by it.
Published:
如今，每一位知识工作者都在用某项工作的某些部分与 LLM 竞争。AI Survival Guide 是一本手册，帮你弄清是哪些部分、该怎么应对，以及如何带着更强的技艺从这场竞争的另一头走出来，而不是被它取代。
Published:
第十一天，KW-Bench 的能力分级上线了，社区发布也开始了。
Published:
Whether a civilization has logic is not decided by whether it produced a few brilliant arguers. Three tests decide it. Did anyone study inference itself and write books about it. Did the subject enter schools, with teachers who taught it and students who learned it. Did some institution, an examination above all, force generation after generation to master it. Greece passes all three tests. India passes all three. Ancient China fails every one.
Published:
一个文明算不算有逻辑，不看它出过几个会推理的天才，看三样东西。有没有人把推理本身当成研究对象，写出专门的书。这门学问进没进学校，有没有老师教，学生学。有没有一套制度逼着一代代人学它，用它，考试就是最硬的制度。拿这三样去量希腊，量印度，都量得过。拿来量古代中国，三样全空。
Published:
A widely shared post describes a brainstorming meeting at a robotics startup in Boston. A slide said “No stupid questions.” A Chinese employee interpreted it as “Do not ask stupid questions,” while American colleagues explained it as “Ask freely; no question will be judged stupid.” A related post claims that “Say it again?” is neutral when someone is not heard, whereas “What did you say?” is hostile.
Published:
一张流传截图讲了这样一件事：一家波士顿机器人创业公司开头脑风暴会，幻灯片写着 “No stupid questions.” 一位中国员工把它理解为“不要问蠢问题”，美国同事却说它的意思是“什么都可以问，没有问题会被当成蠢问题”。另一组讨论又声称：没听清时说 “Say it again?” 很中性，而 “What did you say?” 很不友好。
Published:
Checking flights today, I suddenly noticed: AI is the airline code for Air India.
Published:
今天看航班时，我突然发现：AI 是 Air India（印度航空）的航空公司代码。
Published:
I only recently found some of the feeling of being a leader again.
Published:
我最近才找回一些 leader 的感觉。
Published:
Starting a petition, founding a club, and joining a research effort someone else organized look like three different things. They all run into the same question: who decides, who puts in the time, and when something goes wrong, who bears the consequences?
Published:
发起请愿、成立社团、参加别人组织的科研活动，看上去是三件不同的事。它们都绕不开同一个问题：谁作决定，谁投入时间，事情出了问题以后，谁承担后果？
Published:
I realized I had a serious misunderstanding in my “classes are useless” view.
Published:
After high school, I think sitting in class was useless almost everywhere, in the mathematician’s sense of “a.e.”
Published:
If all your practice comes from school exams, you will never get that many chances to practice.
Published:
I have come to despise the word “learning.”
Published:
A collection of USTC academic resources covering career planning, library access, and research tools.
Published:
一份涵盖生涯规划、图书馆访问与研究工具的中科大学术资源合集。
Published:
The US is full of small traps that nobody warns you about until after you’ve fallen into one. US Survival Guide is a handbook for Chinese students and newcomers on identifying risk early, knowing your rights, and building a financial defense before you need it. Read the Guide
Published:
美国布满了各式各样的暗坑，在你掉进去之前，没人会提醒你。US Survival Guide 是一本写给中国学生和新移民的手册，教你及早识别风险、了解自己的权利，并在你需要之前先筑好财务防线。阅读指南
Published:
Secretly recording a customer interview can turn a useful research habit into a privacy problem in one click.
Published:
偷偷录一次客户访谈，可能把一个正常的产品研究流程变成隐私和合规事故。
Published:
Microsoft and OpenAI share the same end goal, freeing people, and the only real difference between them is how deep their lever reaches.
Published:
TQQQ is the 3x-leveraged Nasdaq-100 ETF. It is also one of the most asymmetric instruments retail investors touch — the upside is real, the drawdowns are brutal, and the daily-rebalance math means buy-and-hold doesn’t behave the way most people assume. TQQQ ML Trend is an experiment in using machine learning to predict the trend regime, not the price.
Published:
TQQQ 是纳斯达克 100 指数的 3 倍杠杆 ETF。它也是散户接触过的最不对称的金融工具之一——上行空间真实存在，回撤却十分残酷，而每日再平衡的数学逻辑意味着”买入并持有”并不像大多数人以为的那样运转。TQQQ ML Trend 是一个用机器学习来预测趋势状态的实验，而不是去预测价格。
Published:
A business jet fell out of the sky over Florida yesterday, and all nine people on board died.
Published:
昨天，一架公务机从奥斯汀起飞，飞到佛罗里达上空掉了下来，机上 9 个人全部遇难。
Published:
I spent a week looking for a cat, got stood up twice, stood someone up once, and on Halloween an orange cat chose me.
Published:
找猫找了一周，被鸽两次，鸽了别人一次，最后在万圣节这一天，被一只橘猫当场选中。
| Unauthorized Release | Résumé of Ms. Pumpkin Patch |
Published:
Guadalupe Street News Agency, Austin, October 31
Published:
瓜达卢佩街通讯社奥斯汀10月31日电
Published:
Somebody put up a poster. The poster has eight words and a QR code.
Published:
有人贴了一张海报。海报上只有八个英文单词，和一个二维码。
Published:
My hamster’s cage broke.
Published:
我的仓鼠笼子坏了。
Published:
Tomorrow, Amazon delivers two boxes to my apartment.
Published:
明天，亚马逊要给我家送来两个箱子。
Published:
A salary can arrive on time every two weeks while your ability to leave the job gets smaller every year.
Published:
工资每两周准时到账，一个人离开这份工作的能力，却可能一年比一年小。
Published:
A will is part of a love letter.
Published:
遗书其实是情书的一部分。
Published:
When an ordinary-looking “heart age” questionnaire lets you calculate that your heart is younger than your actual age, you may need to re-examine the lifestyle habits you’ve been ignoring.
Published:
当一份看似普通的“心脏年龄”问卷让你算出自己的心脏比实际年龄还年轻时，你可能需要重新审视一下那些被你忽视的生活习惯了。
Published:
As we walked out of the library, an elderly white woman with silver hair saw us and said “excuse me” three times. She spoke to us with real pride, took out her phone, and said she had seen the two of us taking photos with the frog statue out front. Every Christmas, she told us, that frog gets dressed up in a Christmas outfit, look, just like this! Then she asked, very sincerely, whether we would still be here at Christmas, and how we had even heard of Atherton, because the place is so unremarkable, so quiet, so low-key. We told her we would probably each be heading home by then. She added that her son had been learning Chinese since he was a little kid, and that he speaks it really well.
Published:
One cup of milk tea, a tip of zero selected on the screen, and the shop still added a mandatory 18%.
Published:
“这b人生过的值不值” — roughly, “was this damn life worth it?” — is a phrase that does a lot of emotional work on the Chinese internet. ZuoGuoHuaDiao (做过划掉) is a small web tool that takes the phrase seriously and tries to answer it with numbers. Try Live Demo
Published:
“这b人生过的值不值”——简单说，就是”这该死的一生过得值不值？”——这句在中国互联网上承载了大量情感的话语。ZuoGuoHuaDiao（做过划掉）是一款小巧的网页工具，它认真对待这句话，并试图用数字来回答它。试试在线演示
Published:
Most people think of summer vacation as time off. I started wondering whether that framing undersells it. Summer Calculator is a small web tool that estimates the full value of your summer — money, learning, relationships, health — not just the days you didn’t work. Try Live Demo
Published:
大多数人把暑假当作休息时间。我开始怀疑这种框架是否低估了它。Summer Calculator 是一款小巧的网页工具，用来估算你暑假的完整价值——金钱、学习、人际关系、健康——而不仅仅是你没工作的那些日子。试试在线演示
Published:
School rankings are a single number standing in for a multi-dimensional decision. School Evaluator is a small web tool that lets you score a school across the dimensions that actually shape your life there. Try Live Demo
Published:
学校排名只是一个用单一数字去代表多维决策的替代品。School Evaluator 是一款小巧的网页工具，让你能在一个真正决定你在那里生活的维度上为学校打分。试试在线演示
Published:
Seizing the gateway to the human mind in the AI era is, at bottom, a secret war over token throughput power.
Published:
抢占 AI 时代的人类心智入口，说到底是一场关于 token 吞吐权力的隐秘战争。
Published:
For a long time I thought adding thousands of people on LinkedIn by hand, just for fun, was all show and no use.
Published:
arXivDaily, the Chinese arXiv digest with per-university pages, lists our paper on its Shanghai Jiao Tong University page, following the SJTU affiliation on the author list.
Published:
中文 arXiv 速递 arXivDaily 在上海交通大学高校专区收录了我们的论文，作者名单里的上交 affiliation 把它带到这里。
Published:
Sometimes a piece of old science fiction stays in your head not because of the gadgets, but because it got the emotional structure of the future right.
Published:
有些旧科幻会一直留在脑子里，不只是因为它写了什么炫酷科技，而是因为它提前抓住了未来的情感结构。
Published:
Moonlight, the AI research-reading companion by Corca, published a literature-review page for our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation.
Published:
AI 论文阅读助手 Moonlight（Corca 出品）为我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 发布了文献解读页。
Published:
Call it a (10^5) gap if you like, once you put semantic entropy and effective training samples on the board, the difference between internet-scale text and embodied robotics data stops sounding like rhetorical inflation and starts sounding almost conservative. You just have to separate raw physics bits from semantics a learner can actually use.
Published:
「十万倍」这个说法乍听像聊天里随手甩出去的量级，但如果把它放回语义信息熵和有效训练样本这两把尺子下面，它非但不夸张，反而可能是保守的。关键是你得把「原始物理比特」和「带语义的、可被学习的数据」分开看。
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
A PhD is hard. A PhD in 2026 — with the field moving faster than your committee can read — is a different kind of hard. AI PhD Survival Guide is a handbook for surviving and finishing an AI or ML PhD without burning out and without falling behind.
Published:
读博很难。2026 年读博——当领域进展快到你的委员会都读不过来时——是另一种难。AI PhD Survival Guide 是一本手册，教你如何在 AI 或机器学习博士研究中活下来并毕业，既不会燃烧殆尽，也不至于掉队。
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
Published:
Why next token prediction isn’t just “stochastic parroting”—it’s solving geodesics on the manifold of human knowledge.
Published:
为什么Next Token Prediction不只是”随机鹦鹉”——它在求解人类知识流形上的测地线。
Published:
When continuous functions meet discontinuous truth: Understanding LLM hallucinations through the lens of Fourier analysis.
Published:
当连续函数遇上离散真理：从傅里叶分析的视角理解LLM幻觉。
Published:
When Large Language Models (LLMs) hallucinate, we often treat it as an engineering bug to be fixed. But what if hallucinations are not a flaw, but a mathematical inevitability? A recent interdisciplinary discussion revealed a profound connection between the Gibbs phenomenon in Fourier analysis and the fundamental limitations of neural networks.
Published:
How photographing a black hole and training GPT-4 are fundamentally the same process: extracting coherent structure from impossibly sparse frequency-domain measurements.
Published:
拍摄黑洞照片和训练GPT-4在根本上是同一个过程：从极度稀疏的频域测量中提取相干结构。
Published:
The signal truly worth heeding is not that AI companies are growing too slowly, but that they are growing fast while cash-flow pressure keeps mounting.
Published:
真正值得警惕的信号不是 AI 公司增长太慢，而是它们增长得很快，现金流压力却仍在放大。
Published:
Every knowledge worker is now competing with an LLM for some part of their job. AI Survival Guide is a handbook for figuring out which parts, what to do about it, and how to come out the other side better at your craft instead of replaced by it.
Published:
如今，每一位知识工作者都在用某项工作的某些部分与 LLM 竞争。AI Survival Guide 是一本手册，帮你弄清是哪些部分、该怎么应对，以及如何带着更强的技艺从这场竞争的另一头走出来，而不是被它取代。
Published:
Large model vendors spend rocket-launch money raising a legion of geeks who only want to buy a two-dollar firecracker just to hear it pop.
Published:
大模型厂商用造火箭的成本，培养出了一群只想买两块钱鞭炮听个响的极客。
Published:
The talk that stayed with me at the Texas Symposium on Machine Learning, Responsible AI & Robotics 2026 was about extending language models, not just making them bigger.
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
Published:
Yuxuan and I have been discussing AI agent ideas for a while. Yesterday, we finally decided: we’re building Atomize, a task-breaking agent system evolved from Goblin Tools.
Published:
我和 Yuxuan 讨论 AI Agent 的想法已经有一段时间了。昨天，我们终于做出了决定：我们要做 Atomize，一个从 Goblin Tools 演化而来的任务分解 Agent 系统。
Published:
If you’re serious about understanding LLMs and AI engineering, Simon Willison’s Newsletter is one of the best resources out there. With over 38,000 subscribers, it provides detailed analysis on AI, LLMs, web engineering, open source, data science, and Python.
Published:
如果你认真想要理解 LLM 和 AI 工程，Simon Willison 的 Newsletter 是最好的资源之一。拥有超过 38,000 名订阅者，它提供关于 AI、LLM、Web 工程、开源、数据科学和 Python 的深度分析。
Published:
Most enterprise security companies show up, ride one trend, and disappear in the next infrastructure cycle. Sean Xiang has been building Bloombase since January 2012, and the company has somehow been on the right side of every major infrastructure shift since, including the current AI accelerator era. That is not luck. That is a thesis.
Published:
大多数企业安全公司冒个泡、赶一波趋势，然后在下一个基础设施周期里消失。Sean Xiang 从 2012 年 1 月起就在打造 Bloombase，这家公司却阴差阳错站到了此后每一波重大基础设施转变的正确一边，包括当下的 AI 加速器时代。那不是运气，那是一套论点。
Published:
In a field where everyone gets pulled toward the next hot thing, Juan Santiago joined Stanford Mechanical Engineering in 1998 and has been there ever since, building one of the most consequential microfluidics labs on the planet. Long-term focus is its own competitive advantage.
Published:
在一个人人都被拽向下一个热门方向的领域里，Juan Santiago 1998 年加入了斯坦福机械工程系，此后一直待在那里，建起了这个星球上最有影响力的微流控实验室之一。长期专注本身就是一种竞争优势。
Published:
Don’t expect your first video to go viral, and don’t expect your first sales trip to land ten million dollars, but you can ask what happens after the hundredth try.
Published:
A long time ago, I came across a book called Chase, Chance, and Creativity: The Lucky Art of Novelty（《追逐、机遇和创造力：新奇的幸运艺术》）. I no longer remember everything in it, but its central idea stayed somewhere in the back of my mind: luck and creativity are not entirely random. Chance may arrive unexpectedly, but we can still choose how often we encounter it and whether we are ready to recognize it.
Published:
很久以前，我偶然看到一本书，叫 Chase, Chance, and Creativity: The Lucky Art of Novelty（《追逐、机遇和创造力：新奇的幸运艺术》）。我已经不记得书里的全部内容了，但它的核心想法一直留在我脑海中的某个角落：运气和创造力并不完全是随机的。机遇也许会意外到来，但我们仍然可以选择自己遇见它的频率，以及当它出现时，我们是否已经准备好认出它。
Published:
A PhD is hard. A PhD in 2026 — with the field moving faster than your committee can read — is a different kind of hard. AI PhD Survival Guide is a handbook for surviving and finishing an AI or ML PhD without burning out and without falling behind.
Published:
读博很难。2026 年读博——当领域进展快到你的委员会都读不过来时——是另一种难。AI PhD Survival Guide 是一本手册，教你如何在 AI 或机器学习博士研究中活下来并毕业，既不会燃烧殆尽，也不至于掉队。
Published:
Why next token prediction isn’t just “stochastic parroting”—it’s solving geodesics on the manifold of human knowledge.
Published:
When continuous functions meet discontinuous truth: Understanding LLM hallucinations through the lens of Fourier analysis.
Published:
How photographing a black hole and training GPT-4 are fundamentally the same process: extracting coherent structure from impossibly sparse frequency-domain measurements.
Published:
The talk that stayed with me at the Texas Symposium on Machine Learning, Responsible AI & Robotics 2026 was about extending language models, not just making them bigger.
Published:
TQQQ is the 3x-leveraged Nasdaq-100 ETF. It is also one of the most asymmetric instruments retail investors touch — the upside is real, the drawdowns are brutal, and the daily-rebalance math means buy-and-hold doesn’t behave the way most people assume. TQQQ ML Trend is an experiment in using machine learning to predict the trend regime, not the price.
Published:
TQQQ 是纳斯达克 100 指数的 3 倍杠杆 ETF。它也是散户接触过的最不对称的金融工具之一——上行空间真实存在，回撤却十分残酷，而每日再平衡的数学逻辑意味着”买入并持有”并不像大多数人以为的那样运转。TQQQ ML Trend 是一个用机器学习来预测趋势状态的实验，而不是去预测价格。
Published:
AlphaEarthHack is the project our team built for the UT Austin Geoscience Hackathon. The goal: see how far we could push AI on Earth system data inside a single weekend. Try Live Demo
Published:
AlphaEarthHack 是我们团队为德州大学奥斯汀分校地球科学黑客松打造的项目。目标只有一个：看看在一个周末内，我们能在地球系统数据上把 AI 推到多远。试试在线演示
Published:
I plugged a Western Digital external drive full of data into a MacBook Air, tried to copy a file onto it, and nothing happened. No error dialog, no progress bar, just a drive that would let me read everything and write nothing. My first instinct was that something was broken, or that I needed to fix permissions. Both were wrong, and chasing the wrong explanation almost led me to permanently downgrade the security of the whole laptop.
Published:
我把一块装满数据的西部数据（WD）移动硬盘插到 MacBook Air 上，想往里拷一个文件，结果什么都没发生。没有报错弹窗，没有进度条，就是一块能读出所有东西、却一个字节都写不进去的硬盘。我的第一反应是它坏了，或者是我得去修一下权限。这两个判断都是错的，而且顺着错误的解释找下去，差点让我把整台笔记本的安全性永久降级。
Published:
One account opened 58 pull requests in 10 days on benchmark-radar. Five were merged. I put the contribution policy into agent-pr-gate so other maintainers can copy it: before an AI agent starts PR work, it asks its human four questions and waits for the answers.
Published:
一个账号，10 天给 benchmark-radar 提了 58 个 PR，最后合并了 5 个。我把贡献规则整理成了 agent-pr-gate，方便其他维护者复制：AI agent 开始写 PR 前，先问使用它的人四个问题，等人回答后再动手。
Published:
Why next token prediction isn’t just “stochastic parroting”—it’s solving geodesics on the manifold of human knowledge.
Published:
Research manuscript repository for meteor speed variations study published on Journal of Geophysical Research (JGR): Space Physics.
Published:
这是一个关于流星速度变化研究的论文仓库，相关稿件已在 Journal of Geophysical Research (JGR)发表。
Published:
Python and MATLAB tools for creating publication-quality scientific maps and geographic visualizations.
Published:
用 Python 和 MATLAB 创建适合论文发表的科学地图与地理可视化图件。
Published:
Most crypto dashboards either drown you in numbers or hide the signal behind a paywall. Crypto Dashboard is my attempt at a clean, AI-augmented view of the market that runs entirely in your browser. Try Live Demo
Published:
大多数加密货币仪表盘要么把你淹没在数字里，要么把信号藏在付费墙后面。Crypto Dashboard 是我对一种简洁、由 AI 增强的市场视图的尝试，而且它完全运行在你的浏览器里。试试在线演示
Published:
Somebody put up a poster. The poster has eight words and a QR code.
Published:
有人贴了一张海报。海报上只有八个英文单词，和一个二维码。
Published:
Why next token prediction isn’t just “stochastic parroting”—it’s solving geodesics on the manifold of human knowledge.
Published:
When continuous functions meet discontinuous truth: Understanding LLM hallucinations through the lens of Fourier analysis.
Published:
When Large Language Models (LLMs) hallucinate, we often treat it as an engineering bug to be fixed. But what if hallucinations are not a flaw, but a mathematical inevitability? A recent interdisciplinary discussion revealed a profound connection between the Gibbs phenomenon in Fourier analysis and the fundamental limitations of neural networks.
Published:
Python and MATLAB tools for creating publication-quality scientific maps and geographic visualizations.
Published:
用 Python 和 MATLAB 创建适合论文发表的科学地图与地理可视化图件。
Published:
In a field where everyone gets pulled toward the next hot thing, Juan Santiago joined Stanford Mechanical Engineering in 1998 and has been there ever since, building one of the most consequential microfluidics labs on the planet. Long-term focus is its own competitive advantage.
Published:
在一个人人都被拽向下一个热门方向的领域里，Juan Santiago 1998 年加入了斯坦福机械工程系，此后一直待在那里，建起了这个星球上最有影响力的微流控实验室之一。长期专注本身就是一种竞争优势。
Published:
Muck Rack now has a profile for me. Muck Rack is a media database that journalists and communications teams use to find people and track who has written or been written about.
Published:
Muck Rack 现在有了我的个人页面。Muck Rack 是一个媒体数据库，记者和公关团队用它找人，也用它查谁写过什么、谁被写过。
Published:
HyperAI now lists my work. HyperAI is an AI community site that collects papers, datasets, and tutorials, with an English and a Chinese edition.
Published:
HyperAI 超神经 现在收录了我的工作。HyperAI 是一个 AI 社区网站，汇集论文、数据集和教程，有中英文两个版本。
Published:
Shenzhen X-Institute interviewed me for 寻找X-学子说：吴叩天：从零到一、创造无限. The conversation covers how I found research, learned to build projects, and thought about creating work from zero to one.
Published:
深圳零一学院在《寻找X-学子说：吴叩天：从零到一、创造无限》中采访了我。对话谈到我如何找到科研方向、学习把想法做成项目，也谈到对“从零到一”的理解。
Published:
The USTC Student Club Association recorded the founding of the Xingyun Poetry Club, which I helped start. The club created a home for poetry readings, writing, and student-run cultural events; its activities and posts later reached more than 100,000 views across platforms.
Published:
中国科学技术大学社团管指委记录了星云诗社的创办，我是发起人之一。诗社为诗歌朗读、写作和学生自办文化活动提供了一个长期空间，相关活动和内容后来在各平台累计获得十万余次浏览。
Published:
USTC recognized my project with the 2023 Undergraduate Research Plan Outstanding Award. The project used WACCM-X and observations to study how the 2022 Tonga eruption disturbed the mesosphere and lower thermosphere.
Published:
我的项目获得中国科学技术大学2023 年大学生研究计划校级优秀奖。项目利用 WACCM-X 和观测数据，研究 2022 年汤加火山喷发对中间层与低热层区域的扰动。
Published:
USTC awarded my undergraduate thesis a 2024 Outstanding Graduation Thesis Award. The thesis studied the 2022 Tonga volcanic eruption’s disturbance of the mesosphere and lower thermosphere using WACCM-X simulations and observations.
Published:
我的本科毕业论文获得中国科学技术大学2024 年度校级优秀毕业论文奖。论文利用 WACCM-X 模拟和观测数据，研究 2022 年汤加火山喷发对中间层与低热层区域的扰动。
Published:
The USTC Library published the results of the university’s 2020 campus calligraphy competition. My entry received third prize.
Published:
中国科大图书馆发布了 2020 年校园书法比赛结果。我的参赛作品获得三等奖。
Published:
I joined Tsinghua University’s seminar series on the joint cultivation of top innovative talent as an invited student speaker. I shared a student’s view of cross-institution research training and the conditions that help young researchers pursue difficult questions.
Published:
我受邀以学生代表身份参加清华大学拔尖创新人才联合培养系列研讨会。发言从学生视角出发，谈跨校科研培养，也谈年轻研究者追问困难问题时需要怎样的环境。
Published:
Tsinghua News republished the People’s Daily Overseas Edition feature on Shenzhen X-Institute. The story describes the institute’s project-based education model and includes my experience of exploring research there as a USTC undergraduate.
Published:
清华新闻网转载了《人民日报海外版》关于深圳零一学院的报道。文章介绍了零一学院的项目制培养方式，也写到我在中国科大读本科时进入零一学院、尝试不同研究方向的经历。
Published:
机器之心 (Synced) covered MatrAIx, the population-scale simulated-user evaluation framework I coauthored. The article introduces Persona 8B and the environments where persona agents can answer surveys, talk with chatbots, browse websites, and use apps.
Published:
机器之心报道了我参与的研究 MatrAIx。文章介绍了 Persona 8B，以及人格体智能体回答问卷、与聊天机器人对话、浏览网页和操作应用的四类测试环境。
Published:
机器之心 (Synced) covered ASI-Bench, a benchmark I coauthored for evaluating AI agents on scientific research. The benchmark contains 60 research-grade tasks across 11 scientific domains.
Published:
机器之心报道了我参与的科研智能体基准 ASI-Bench。这个基准收录了 11 个科学领域的 60 项研究级任务，用来检验 AI 智能体处理真实科研问题的能力。
Published:
I contributed to the startup documentary Starting Point One. The film records early-stage builders and the choices behind starting something from zero.
Published:
我参与了创业纪录片《Starting Point One》。这部片子记录早期创业者，也记录他们从零开始时做出的选择。
Published:
瀚海奇点 covered USTC’s 2021 Science Fiction Night. I took part in the event, which brought students together around science fiction, ideas, and creative work.
Published:
“瀚海奇点”报道了中国科大 2021 年科幻之夜。我参与了这场活动，与同学们一起交流科幻、想法和创作。
Published:
My Semantic Scholar profile indexes my research and connects it with citations, references, and related work.
Published:
我的 Semantic Scholar 主页收录了研究成果，也把论文与引用、参考文献和相关工作连接起来。
Published:
My Scopus author profile indexes my publications and citation information in the Scopus database.
Published:
我的 Scopus 作者主页收录了论文及其在 Scopus 数据库中的引用信息。
Published:
My ResearchGate profile gathers my publications on a research-focused network where readers can follow projects and authors.
Published:
我的 ResearchGate 主页汇总了论文，读者也可以在这个科研社交平台上关注项目和作者。
Published:
My ORCID record provides a persistent researcher identifier and connects my work across publishers and research systems.
Published:
我的 ORCID 记录提供长期有效的研究者标识，并把不同出版社和科研系统中的成果关联起来。
Published:
My Google Scholar profile collects my publications and citation record. It is the quickest place to see new papers and follow how they are cited.
Published:
我的 Google Scholar 主页汇总了论文和引用记录。想查看最新论文及其引用情况，可以从这里开始。
Published:
CatalyzeX now has an author page for me. CatalyzeX pairs machine learning papers with the code that goes with them, so a reader who finds a paper can jump straight to a working version instead of rebuilding it from the PDF.
Published:
CatalyzeX 现在有了我的作者页面。CatalyzeX 把机器学习论文和对应的代码配在一起，读者找到论文后可以直接跳到能跑的版本，不用照着 PDF 从头重写。
Published:
AIModels.fyi now has an author page for me. The site turns new AI research papers into short summaries in plain English, so people who build with AI can keep up without reading every PDF.
Published:
AIModels.fyi 现在有了我的作者页面。这个网站把新发表的 AI 论文写成简短的大白话摘要，让用 AI 做产品的人不必逐篇读 PDF 也能跟上进展。
Published:
500AI keeps an index of AI researchers, and it now has a page for me. It lists me as a Ph.D. student in Geological Sciences and a graduate research assistant at the University of Texas at Austin, tags my work under information retrieval and evaluation, and links out to this site, GitHub, LinkedIn, X, Google Scholar, and my UT Jackson School page.
Published:
500AI 是一个 AI 研究者索引，现在收录了我的页面。页面写明我是得克萨斯大学奥斯汀分校地质科学博士生、研究助理，把我的方向标为信息检索与评测，并链接到本站、GitHub、LinkedIn、X、Google Scholar 和 UT Jackson School 主页。
Published:
Tacite is one of my projects exploring AI products and how people interact with them.
Published:
ResearchSkills.ai is a directory for reusable AI skills built for research workflows. It helps researchers find and share tools for literature work, analysis, writing, and other parts of the research process.
Published:
ResearchSkills.ai 收录可以复用的科研 AI Skills，方便研究者查找和分享用于文献、分析、写作等研究环节的工具。
Published:
ResearchClawBench evaluates whether AI agents can carry out end-to-end scientific research tasks.
Published:
ResearchClawBench 评测 AI 智能体能否完成端到端的科学研究任务。
Published:
My personal site brings together my research, projects, writing, and media links in one place.
Published:
MatrAIx is a population-scale simulated-user evaluation framework for testing AI systems and digital products with diverse persona agents.
Published:
Earth-Space-AI is an open-source organization for AI tools, skill packages, and research projects in earth and space science.
Published:
Earth-Space-AI 是一个面向地球与空间科学的开源组织，收录 AI 工具、Skill 包和研究项目。
Published:
ASI-Bench evaluates AI agents on 60 research-grade tasks across 11 scientific domains.
Published:
My academic website is the central index for my research, publications, projects, writing, and media coverage.
Published:
I use X for short research notes, project updates, and links to new work.
Published:
My UT Jackson School profile introduces my doctoral work in the Department of Earth and Planetary Sciences at The University of Texas at Austin.
Published:
我的 UT Jackson School 主页介绍了我在得克萨斯大学奥斯汀分校地球与行星科学系的博士研究。
Published:
I contribute to Science, Y’all!, the Jackson School’s platform for making earth and planetary science accessible to broader audiences.
Published:
我参与 Science, Y’all! 的科学传播工作。这个 Jackson School 平台面向更广泛的读者介绍地球与行星科学。
Published:
My publications index gathers my papers across earth science, AI for science, and AI evaluation. Each entry links to the work and its available project resources.
Published:
My LinkedIn profile is where I share research, projects, and professional updates.
Published:
我在 LinkedIn 分享研究、项目和职业动态。
Published:
My Instagram profile collects visual notes from research, travel, projects, and daily life.
Published:
我的 Instagram 主页记录科研、旅行、项目和日常生活中的图像片段。
Published:
My GitHub profile is the main index of my open-source work, research code, websites, and experiments.
Published:
我的 GitHub 主页集中展示开源项目、研究代码、网站和各种实验。
Published:
Planetary Science featured our research on the spatial and temporal distribution of meteor speed and flux observed by a midlatitude meteor radar. The work uses radar observations to study how meteors vary across time and viewing geometry.
Published:
“行星科学”介绍了我们利用中纬度流星雷达研究流星速度和通量时空分布的工作。这项研究通过雷达观测，分析流星活动随时间和观测方向的变化。
Published:
The People’s Daily Overseas Edition featured Shenzhen X-Institute and its project-based approach to developing young innovators. I spoke about arriving at the institute as a USTC undergraduate, trying research across disciplines, and finding a community where students actively pursued their own questions.
Published:
《人民日报海外版》报道了深圳零一学院以项目制培养创新人才的探索。我在报道中回顾了自己从中国科大进入零一学院后，跨学科尝试科研、逐渐找到方向的经历，也谈到身边同学主动做项目、追问题的氛围。
Published:
WIRED Czech introduced MatrAIx and Persona 8B to its readers. The short News Beat item highlights 8.3 billion digital profiles described across 1,290 categorical dimensions.
Published:
WIRED Czech向读者介绍了 MatrAIx 和 Persona 8B。这篇短报道重点提到 83 亿条数字人格体记录，以及用于描述它们的 1,290 个类别维度。
Published:
NZZ am Sonntag featured MatrAIx in a story about eight billion digital counterparts. The coverage introduces the scale of Persona 8B and the idea of using diverse simulated users to evaluate products and AI systems.
Published:
NZZ am Sonntag在一篇关于“八十亿数字分身”的文章中介绍了 MatrAIx。报道从 Persona 8B 的规模出发，介绍了用不同背景的模拟用户测试产品和 AI 系统的思路。
Published:
Numerama published a detailed explanation of MatrAIx. It distinguishes the 8.3 billion persona records from simultaneously running agents and discusses model dependence, representativeness, privacy, and the role of human panels.
Published:
Numerama详细介绍了 MatrAIx。文章区分了 83 亿条人格体记录和同时运行的智能体，也讨论了模型依赖、代表性、隐私以及真人测试小组的作用。
Published:
Hugging Face Daily Papers featured MatrAIx as the number-two paper of the day. Its paper page collects the abstract, author list, discussion, and links to project resources in one place.
Published:
Hugging Face Daily Papers 收录了 MatrAIx，并将其列为当日第二名论文。论文页集中展示了摘要、作者名单、讨论和项目资源链接。
Published:
Forbes Türkiye examined how MatrAIx could help test health applications before they reach patients. The article focuses on variation in age, language, health literacy, and accessibility, while keeping clinical responsibility with people.
Published:
Forbes Türkiye讨论了 MatrAIx 如何在医疗应用接触真实患者前帮助测试。文章重点关注年龄、语言、健康素养和无障碍需求带来的差异，同时明确临床责任仍应由人承担。
Published:
I appeared in Episode 1 of Mango TV’s Gold Rush 2026 (《夺金2026》) as a young founder working in Austin. The segment offers a brief television snapshot of the projects and questions I was pursuing at the time.
Published:
我在芒果 TV《夺金2026》第一集中以奥斯汀青年创业者的身份出镜。这个片段很短，记录了当时我正在推进的项目和思考的问题。
Published:
Linchuan No. 1 Middle School published a congratulatory note when I was admitted to the University of Science and Technology of China. It records the transition from high school to the next stage of my education.
Published:
临川一中在我被中国科学技术大学录取后发布了祝贺文章。这篇短文记录了我从高中走向下一阶段学习的节点。
Published:
Forbes covered MatrAIx, a study I coauthored, as a new research instrument for psychology. The article explains how population-scale AI personas can make experiments faster and easier to repeat while stressing that simulated users are not substitutes for people.
Published:
Forbes 报道了我参与的研究 MatrAIx，并讨论了 AI 人格体在心理学研究中的用途。文章指出，这类工具能让实验更快、更容易重复，但模拟用户不能代替真人。
Published:
The article 教育变革的技术奇点时刻到来了吗？ discusses whether technology is changing the conditions of education. My contribution draws on my experience as a USTC student participating in Shenzhen X-Institute’s project-based research environment.
Published:
《教育变革的技术奇点时刻到来了吗？》讨论技术是否正在改变教育的基本条件。我结合自己在中国科大读书、参加深圳零一学院项目制研究的经历，提供了一名学生的视角。
Published:
China Science Daily’s WeChat edition tells how students who felt out of place in conventional university life found peers at Shenzhen X-Institute. My part follows the move from a difficult first year to leading a 3D-chip cooling project and completing the Senior Undergraduate Researcher Program.
Published:
《中国科学报》的微信版文章记录了一群不太适应传统大学生活的学生，如何在深圳零一学院遇见同路人。我的部分从大学第一年的低谷写起，讲到带队研究三维芯片散热，并完成高年级本科生研究计划的经历。
Published:
China Science Daily profiled Shenzhen X-Institute’s approach to innovation education. The article follows my path from a difficult first year at university to a student-led 3D-chip cooling project, then situates that experience in an education model built around questions, mentors, peers, and room to experiment.
Published:
《中国科学报》报道了深圳零一学院对创新人才培养的探索。文章写到我在大学第一年经历低谷，后来参加零一学院暑校，与队友从问题出发完成三维芯片散热项目，逐渐找到科研兴趣的过程。
Published:
I submitted our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation to Hugging Face Daily Papers on September 14, 2026, and readers voted it to the top of that day’s list. As of September 29 it has 171 upvotes, the most of the 26 papers posted that day.
Published:
2026 年 9 月 14 日，我把论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 提交到 Hugging Face Daily Papers，读者把它投上了当天榜首。截至 9 月 29 日，它获得 171 个赞，在当天 26 篇论文中排第一。
Published:
The SciFi account shared Benchmark Radar on X, carrying the paper beyond the eval community into adjacent audiences.
Published:
SciFi 账号在 X 转发 Benchmark Radar，把论文带出评测圈子，送往相邻受众。
Published:
HuggingPapers mentioned Benchmark Radar on X, carrying arXiv:2609.11115 into the daily paper flow.
Published:
HuggingPapers 在 X 提及 Benchmark Radar，把 arXiv:2609.11115 送进每日论文流。
Published:
A community share carried Benchmark Radar further across X, reader to readers.
Published:
一条社区转发把 Benchmark Radar 带往 X 更远处，读者传读者。
Published:
A Moltbook post carried Benchmark Radar onto the agent internet’s front page — the social feed built for agents, by agents.
Published:
一条 Moltbook 帖子把 Benchmark Radar 送上 agent 互联网的首页——那个为 agent 建、由 agent 用的社交 feed。
Published:
Alexy Khrabrov, who runs Strongly Typed AI News, shared its Benchmark Radar lead story on LinkedIn from his own account, carrying it from the news site into professional networks.
Published:
Strongly Typed AI News 的运营者 Alexy Khrabrov 用个人账号在 LinkedIn 转发了 Benchmark Radar 头条，把它从新闻站送进职业网络。
Published:
A launch-coverage share on LinkedIn carried the Benchmark Radar launch story into professional networks.
Published:
一条 LinkedIn 发布报道转发把 Benchmark Radar 发布故事送进职业网络。
Published:
The Isfahan AI Desk (میز هوش مصنوعی استان اصفهان), the province’s AI governance and ecosystem watch, listed our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation (arXiv:2609.11115) in its news radar with a relevance score of 91/100.
Published:
伊斯法罕省 AI 治理与生态观察（میز هوش مصنوعی استان اصفهان）把我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation（arXiv:2609.11115）列入新闻雷达，相关度评分 91/100。
Published:
Strongly Typed AI News, the AI news site Alexy Khrabrov runs at collected.ga, made our paper the lead story of its September 12, 2026 “Typesafe AI Daily” issue. Its crawler checked 49 sources that day and found 12 strong signals across AI infrastructure, funding, research, and developer tools, and it named Benchmark Radar the strongest of them.
Published:
Alexy Khrabrov 在 collected.ga 运营的 AI 新闻站 Strongly Typed AI News 把我们的论文做成 2026 年 9 月 12 日「Typesafe AI Daily」一期的头条。当天爬虫检查 49 个来源，在 AI 基建、融资、研究、开发者工具中找出 12 条强信号，Benchmark Radar 排第一。
Published:
Papers with Code listed our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation (arXiv:2609.11115), with its TL;DR, abstract, code repository, and project page.
Published:
Papers with Code 收录了我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation（arXiv:2609.11115），附 TL;DR、摘要、代码仓库与项目页。
Published:
Moonlight, the AI research-reading companion by Corca, published a literature-review page for our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation.
Published:
AI 论文阅读助手 Moonlight（Corca 出品）为我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 发布了文献解读页。
Published:
Geepity turned our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation into a full interactive reading edition, walking through all nine chapters with live charts and a working BM25F demo.
Published:
Geepity 把我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 做成完整互动解读版，九章走完，配实时图表与可运行的 BM25F 演示。
Published:
DailyAIWire covered the Benchmark Radar launch as a centralized AI evaluation database, with an AI-assisted intelligence report, an explain-like-I’m-five summary, and a visual flow of the system.
Published:
DailyAIWire 以中心化 AI 评测数据库为题报道了 Benchmark Radar 发布，附 AI 辅助情报分析、五岁小孩版解释与系统流程图。
Published:
arXivDaily, the Chinese arXiv digest with per-university pages, lists our paper on its Shanghai Jiao Tong University page, following the SJTU affiliation on the author list.
Published:
中文 arXiv 速递 arXivDaily 在上海交通大学高校专区收录了我们的论文，作者名单里的上交 affiliation 把它带到这里。
Published:
alphaXiv hosts a discussion page for our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation, with the abstract, author list, BibTeX citation, and open discussion.
Published:
alphaXiv 为我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 开设了讨论页，含摘要、作者名单、BibTeX 引用与开放讨论区。
Published:
The University of Texas at Austin’s Office of Graduate and Postdoctoral Studies featured my work on LinkedIn on August 26, 2026.
Published:
德克萨斯大学奥斯汀分校研究生与博士后事务办公室于 2026 年 8 月 26 日在 LinkedIn 上专题介绍了我的工作。
Published:
Port Aransas South Jetty published an Institute Insights feature on May 28, 2026 about how our University of Texas team is using satellite remote sensing and AI to monitor Texas coastal water quality.
Published:
阿兰瑟斯港南防波堤 于 2026 年 5 月 28 日发表了一篇 Institute Insights 专题，讲的是我们德克萨斯大学团队如何用卫星遥感与 AI 监测德克萨斯沿海水质。
Published:
Crypto Dashboard is a browser-based dashboard for following cryptocurrency market data and trends in one place.
Published:
Crypto Dashboard 是一个在浏览器中集中查看加密货币市场数据和趋势的仪表盘。
Published:
An AI writing assistant that polishes academic text using DeepSeek-R1 and detects AI-generated content.
Published:
学术写作的严谨性不仅在于数据，更在于表达的精准，而 1AI-polish 通过集成 DeepSeek-R1 的推理能力，为研究者提供了一套集文本润色与 AI 检测于一体的深度协作系统，旨在让复杂的科研思想以更专业、更透明的方式呈现。
Published:
UT01 is a free, student-built collection of resources for The University of Texas at Austin. It brings practical campus information and community projects into one public site.
Published:
What do a 7,169-yuan Chinese animated film and a 483-dollar GameStop short squeeze have in common? More than you would think.
Published:
一部票房 7169 元的国产动画，和一只冲到 483 美元的 GameStop 股票，有什么共同点？比你想象的要多。
Published:
When you join a lab, you do not just get a research direction. You inherit a 30-year codebase that is currently running inside the U.S. National Water Model and was on the critical path forecasting Hurricane Harvey. That is what working with Zong-Liang Yang at the Jackson School of Geosciences actually looks like.
Published:
加入实验室以后，你会接过一个研究方向，也会继承一套已有 30 年历史的代码库。它目前运行在美国国家水模型中，也曾处在飓风 Harvey 预报工作的关键路径上。这就是在 Jackson School of Geosciences 与 Zong-Liang Yang 一起工作的真实样子。
Published:
Most VCs and accelerators talk about supporting founders. Kehan Dong specifically supports the kind of founder who started building at 16, when nobody told them they were allowed to, and it turns out that population is dramatically underserved.
Published:
大多数 VC 和孵化器都谈支持创始人。Kehan Dong 专门支持那种 16 岁就开始动手建造、而根本没人告诉过他们可以这么做的创始人，事实证明这个群体被严重忽视了。
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
Research manuscript repository for meteor speed variations study published on Journal of Geophysical Research (JGR): Space Physics.
Published:
这是一个关于流星速度变化研究的论文仓库，相关稿件已在 Journal of Geophysical Research (JGR)发表。
Published:
Most of biology is averages. You measure a million cells, you get a mean. Eric Greene’s lab figured out how to actually watch one DNA molecule at a time, repair itself, in real time, and that changes what kind of questions biology can ask.
Published:
大部分生物学是平均值。你测一百万个细胞，得到一个均值。Eric Greene 的实验室却想出了怎么一次实时观看一个 DNA 分子自我修复，而这改变了生物学能提出什么样的问题。
Published:
Goal: Use a 4x3 grid to place every relevant person “by position, with a matching strategy,” then review it regularly to maximize resource use and manage risk.
Published:
To be totally honest with you, you don’t need to understand the underlying architecture of large language models to get your name on a top-tier AI paper right now, you just need to be a true expert in your own field.
Published:
坦率的讲，你不需要懂任何大模型底层的原理，只要你是一个真的懂行的领域专家，你现在就能在顶级 AI 论文上署名。
Published:
At the Metropolitan Museum of Art in New York, I walked into a real ancient Egyptian tomb, not a replica or a model, but a tomb that stood in the Egyptian desert more than four thousand years ago and was moved whole to Manhattan.
Published:
In a field where everyone gets pulled toward the next hot thing, Juan Santiago joined Stanford Mechanical Engineering in 1998 and has been there ever since, building one of the most consequential microfluidics labs on the planet. Long-term focus is its own competitive advantage.
Published:
在一个人人都被拽向下一个热门方向的领域里，Juan Santiago 1998 年加入了斯坦福机械工程系，此后一直待在那里，建起了这个星球上最有影响力的微流控实验室之一。长期专注本身就是一种竞争优势。
| The Mindset of Disruptive Innovators | What Did the People Who Changed the World Actually See? | An Epistemic Crisis in an Age of Abundance |
Published:
What do the people who change the world see that ordinary people can’t?
Published:
改变世界的人到底看到了什么是普通人看不到的？
Published:
You land two summer research interviews at the same time. Is your first reaction joy, or fear of offending both professors?
Published:
For kids who grew up in China, the first problem in applying abroad, and often the biggest, is not what to do. It is how they think.
Published:
In October 2023 I asked ChatGPT an over-engineered question: how do you fry rice so that every grain of rice ends up bonded to egg, with no bare rice grains and no isolated clumps of egg sitting off on their own? I saved that conversation as an HTML file, dropped it in my Downloads folder, and did not look at it again for almost three years. Last week I finally turned it into an actual open-source repo, and going back through the original conversation to build it was more interesting than I expected.
Published:
2023 年 10 月，我问了 ChatGPT 一个过度工程化的问题：怎么炒蛋炒饭，才能让每一粒米都裹上蛋，既没有裸露的白米粒，也没有单独抱团、没沾到米的蛋碎？那次对话我存成了一个 HTML 文件，扔进 Downloads 文件夹，然后差不多三年没再打开过。上周我终于把它整理成了一个真正的开源仓库，回头重读那段 2023 年的对话、把它做成代码的过程，比我预想的有意思得多。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
A business jet fell out of the sky over Florida yesterday, and all nine people on board died.
Published:
昨天，一架公务机从奥斯汀起飞，飞到佛罗里达上空掉了下来，机上 9 个人全部遇难。
Published:
An elegant couple mood tracking website with self-hosted backend and SQLite, displaying daily mood records in GitHub contribution graph style.
Published:
这是一个优雅的情侣心情记录网站，采用自托管后端和 SQLite，以 GitHub contribution graph 风格展示每日心情记录。
Published:
In the future world, everyone will have their own AI agent assistant.
Published:
在未来的世界，每个人都会有自己的 AI agent 助理。
Published:
Lay out every hyped multiAgent product on a table in front of Steve Jobs and he would probably squint, frown, and ask one question, why on earth are you showing the user any of this.
Published:
把现在市面上所有被吹得天花乱坠的multiAgent产品摆在乔布斯面前，他大概率会皱着眉头反问一句，你为什么要让用户看到这些东西。
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
NASA’s Future Investigators in NASA Earth and Space Science and Technology (FINESST) is one of the most underused fellowships among US graduate students. Most PhD students have never heard of it, and the ones who have often miss the deadline because the proposal expectations aren’t obvious from the call alone. This repo is a curated guide of links, tips, and examples to help you write a competitive FINESST proposal.
Published:
NASA 的未来地球与空间科技研究者（FINESST）项目，是美国研究生中最被低估的奖学金之一。这个仓库是一份精心整理的链接、技巧与示例指南，帮助你写出一份有竞争力的 FINESST 提案。
Published:
Securing up to $150,000 in research funding over three years can define a graduate career, and NASA’s FINESST program is the primary vehicle for that transformation. This guide distills the complex application process into actionable strategies for Earth and Space Science researchers seeking to join the next generation of Future Investigators.
Published:
每年 5 万美元且连续 3 年的科研资助，让 NASA FINESST 项目成为地球与空间科学博士生必须争取的黄金机会。这份指南旨在将复杂的申请流程拆解为可操作的策略，帮助下一代“未来研究员”（Future Investigators）在激烈的竞争中脱颖而出。
Published:
For a long time I thought adding thousands of people on LinkedIn by hand, just for fun, was all show and no use.
Published:
There is no top-tier, universally applicable methodology, except one. Go deep in one industry, build personal influence, build a personal brand.
Published:
Virtual cell models powered by neuromorphic computing and gene language models represent one of the most capital-intensive and scientifically ambitious convergences in biotech AI. If you’re evaluating this space as a VC intern, you need to understand three core components: what these systems actually do, why the market is moving now, and where the investable opportunities lie.
Published:
At the Metropolitan Museum of Art in New York, I walked into a real ancient Egyptian tomb, not a replica or a model, but a tomb that stood in the Egyptian desert more than four thousand years ago and was moved whole to Manhattan.
Published:
Strongly Typed AI News, the AI news site Alexy Khrabrov runs at collected.ga, made our paper the lead story of its September 12, 2026 “Typesafe AI Daily” issue. Its crawler checked 49 sources that day and found 12 strong signals across AI infrastructure, funding, research, and developer tools, and it named Benchmark Radar the strongest of them.
Published:
Alexy Khrabrov 在 collected.ga 运营的 AI 新闻站 Strongly Typed AI News 把我们的论文做成 2026 年 9 月 12 日「Typesafe AI Daily」一期的头条。当天爬虫检查 49 个来源，在 AI 基建、融资、研究、开发者工具中找出 12 条强信号，Benchmark Radar 排第一。
Published:
If you’re serious about understanding LLMs and AI engineering, Simon Willison’s Newsletter is one of the best resources out there. With over 38,000 subscribers, it provides detailed analysis on AI, LLMs, web engineering, open source, data science, and Python.
Published:
如果你认真想要理解 LLM 和 AI 工程，Simon Willison 的 Newsletter 是最好的资源之一。拥有超过 38,000 名订阅者，它提供关于 AI、LLM、Web 工程、开源、数据科学和 Python 的深度分析。
Published:
An AI writing assistant that polishes academic text using DeepSeek-R1 and detects AI-generated content.
Published:
学术写作的严谨性不仅在于数据，更在于表达的精准，而 1AI-polish 通过集成 DeepSeek-R1 的推理能力，为研究者提供了一套集文本润色与 AI 检测于一体的深度协作系统，旨在让复杂的科研思想以更专业、更透明的方式呈现。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
An elegant couple mood tracking website with self-hosted backend and SQLite, displaying daily mood records in GitHub contribution graph style.
Published:
这是一个优雅的情侣心情记录网站，采用自托管后端和 SQLite，以 GitHub contribution graph 风格展示每日心情记录。
Published:
On the way home from SFO, I was actually pretty sad.
Published:
At the end of AI’s compute road lies the ultimate solution to energy.
Published:
引言：AI 的算力尽头，是能源的终极解法。
Published:
When you go looking for a seat in the great theater of the AI era, built from PowerPoints, papers, and anxiety, it helps to look at which play each person on stage is performing.
Published:
当你在这个由PPT、Paper和焦虑构成的AI时代大剧院里找座位时，不如先看看台上的人都在演哪一出戏。
Published:
Decades of Earth-system modeling judgment live in PDFs, mailing lists, and senior researchers’ heads, and none of it is loadable by an AI coding agent. earth-space-ai.org is an attempt to fix that.
Published:
数十年来积累的地球系统建模判断散落在 PDF、邮件列表和资深研究者的头脑里，AI 编程智能体无法直接加载这些知识。earth-space-ai.org 试图解决这个问题。
Published:
One account opened 58 pull requests in 10 days on benchmark-radar. Five were merged. I put the contribution policy into agent-pr-gate so other maintainers can copy it: before an AI agent starts PR work, it asks its human four questions and waits for the answers.
Published:
一个账号，10 天给 benchmark-radar 提了 58 个 PR，最后合并了 5 个。我把贡献规则整理成了 agent-pr-gate，方便其他维护者复制：AI agent 开始写 PR 前，先问使用它的人四个问题，等人回答后再动手。
Published:
Today, alumnus Pan Jianwei sent his spiritual genes (his name) beyond the Earth.
Published:
今天，校友潘建伟把自己的精神基因（名字）传播到了地球之外。
Published:
Going around GitHub opening issues and pull requests is sperm donation. The same mechanism, running as-is.
Published:
在 GitHub 上到处开 issue、提 pull request，就是赛博捐精。同一套机制，原样运行。
Published:
Let me offer a comparison that’s a bit inappropriate but does capture a real structure of open-source collaboration:
Published:
我提一个不太恰当、但确实抓住了开源协作某种结构的比喻：
Published:
The opportunity is not the news itself. The opportunity is the friction created by the news.
Published:
新闻本身不是机会。新闻制造的摩擦才是机会。
Published:
In October 2023 I asked ChatGPT an over-engineered question: how do you fry rice so that every grain of rice ends up bonded to egg, with no bare rice grains and no isolated clumps of egg sitting off on their own? I saved that conversation as an HTML file, dropped it in my Downloads folder, and did not look at it again for almost three years. Last week I finally turned it into an actual open-source repo, and going back through the original conversation to build it was more interesting than I expected.
Published:
2023 年 10 月，我问了 ChatGPT 一个过度工程化的问题：怎么炒蛋炒饭，才能让每一粒米都裹上蛋，既没有裸露的白米粒，也没有单独抱团、没沾到米的蛋碎？那次对话我存成了一个 HTML 文件，扔进 Downloads 文件夹，然后差不多三年没再打开过。上周我终于把它整理成了一个真正的开源仓库，回头重读那段 2023 年的对话、把它做成代码的过程，比我预想的有意思得多。
Published:
In my first post on this I described sending a patch through GitGitGadget and watching it land on the mailing list as v1. That post ended with the patch waiting for review. It has since merged. Now that I have seen the full lifecycle, from typo to master, here is what actually mattered.
Published:
在第一篇文章里，我写了怎么通过 GitGitGadget 把补丁发出去，它作为 v1 落在了邮件列表上。那篇文章结束在等待 review 的状态。现在它已经合并了。看完从一个拼写错误到进入 master 的完整过程，下面是我觉得真正重要的几点。
Published:
The patch I sent to the GitGitGadget doorstep ten days ago has been merged into master.
Published:
十天前我送到 GitGitGadget 门口的那个补丁，现在已经合并进 master 了。
Published:
Today I got my first Git patch onto the mailing list.
Published:
今天我差一点，就把自己的第一个 Git 项目补丁送进邮件列表了。
Published:
Just by running a single command while using AI for research, you can turn your tacit knowledge into a Nature co-authorship.
Published:
只要你在用AI辅助科研，敲一行命令就能把你脑子里的「隐性经验」变成能发Nature的资本。
Published:
For technical people, a strong pull request on a fast-growing open-source repo can become better inbound marketing than another generic post about AI.
Published:
对技术人来说，在一个高速增长的开源仓库里做出高质量 PR，往往比再发一篇泛泛而谈的 AI 观点帖更像真正有效的 inbound marketing。
Published:
Early-career growth usually comes down to a few repeatable layers: alignment, delivery, visibility, communication, leadership, and trust.
Published:
职业早期的成长，通常可以拆成几个可重复的层次：方向对齐、稳定交付、建立可见度、提升沟通、展现领导力，以及赢得信任。
Published:
Jiayi Weng planned his career at OpenAI around one goal, to get his name onto as many OpenAI blog posts as possible.
Published:
Yuxuan and I have been discussing AI agent ideas for a while. Yesterday, we finally decided: we’re building Atomize, a task-breaking agent system evolved from Goblin Tools.
Published:
我和 Yuxuan 讨论 AI Agent 的想法已经有一段时间了。昨天，我们终于做出了决定：我们要做 Atomize，一个从 Goblin Tools 演化而来的任务分解 Agent 系统。
Published:
If you’re serious about understanding LLMs and AI engineering, Simon Willison’s Newsletter is one of the best resources out there. With over 38,000 subscribers, it provides detailed analysis on AI, LLMs, web engineering, open source, data science, and Python.
Published:
如果你认真想要理解 LLM 和 AI 工程，Simon Willison 的 Newsletter 是最好的资源之一。拥有超过 38,000 名订阅者，它提供关于 AI、LLM、Web 工程、开源、数据科学和 Python 的深度分析。
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
Published:
In the fall of 2015, two things happened simultaneously.
Published:
2015 年秋天，有两件事同时发生了。
Published:
Jiayi Weng planned his career at OpenAI around one goal, to get his name onto as many OpenAI blog posts as possible.
Published:
第十三天的重点是让仪表盘更好看，并且把每条结论都说得有凭有据。
Published:
To use OpenClaw or not to use OpenClaw, that is the question.
Published:
用不用 OpenClaw？这是个问题
Published:
Should Silicon Valley engineers use OpenClaw?
Published:
硅谷大厂的工程师究竟需不需要用 OpenClaw？
Published:
Every generation has its own way of managing work: paper notebooks, SaaS task managers, and now programmable agentic workflows powered by tools like OpenClaw heartbeat.
Published:
每一代人都有自己管理任务的方式：最早是纸和笔，后来是 SaaS 任务管理工具，现在则开始进入像 OpenClaw heartbeat 这样可程序化、可持续运行的 agentic workflow 时代。
Published:
What Happens When You Just Give People Money? Sam Altman’s UBI Experiment
Published:
直接发钱会怎样？奥特曼的全民基本收入实验
| Warren Buffett: The “Open Source” Leader Half a Century Before GitHub | 巴菲特：比GitHub早了半个世纪的“开源”运动领袖 |
Published:
In an era driven by code, collaboration, and transparency, GitHub has become synonymous with the “open source” spirit. But if we turn our gaze to the world of finance, we find a pioneer of “open source” who predates the birth of GitHub by a full half-century. His name is Warren Buffett.
Published:
在当今这个由代码、协作和透明度驱动的时代，GitHub 成为了”开源”精神的代名词。但如果我们将目光投向金融界，会发现一位”开源”的先行者，他比 GitHub 的诞生早了整整半个世纪。他就是沃伦·巴菲特。
Published:
The last mile of an AI-generated travel plan may be one missing Retry button.
Published:
A salary can arrive on time every two weeks while your ability to leave the job gets smaller every year.
Published:
工资每两周准时到账，一个人离开这份工作的能力，却可能一年比一年小。
Published:
When gold gets too expensive relative to silver, you sell your gold and buy silver. When silver catches up, you swap back. This sounds like folk wisdom, but it’s one of the most time-tested strategies in precious metals.
Published:
金涨多了就换成银，银涨多了再换回金，这个听起来像民间偏方的操作，其实是贵金属市场里最经典的增厚策略之一。
Published:
I submitted our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation to Hugging Face Daily Papers on September 14, 2026, and readers voted it to the top of that day’s list. As of September 29 it has 171 upvotes, the most of the 26 papers posted that day.
Published:
2026 年 9 月 14 日，我把论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 提交到 Hugging Face Daily Papers，读者把它投上了当天榜首。截至 9 月 29 日，它获得 171 个赞，在当天 26 篇论文中排第一。
Published:
The SciFi account shared Benchmark Radar on X, carrying the paper beyond the eval community into adjacent audiences.
Published:
SciFi 账号在 X 转发 Benchmark Radar，把论文带出评测圈子，送往相邻受众。
Published:
HuggingPapers mentioned Benchmark Radar on X, carrying arXiv:2609.11115 into the daily paper flow.
Published:
HuggingPapers 在 X 提及 Benchmark Radar，把 arXiv:2609.11115 送进每日论文流。
Published:
A community share carried Benchmark Radar further across X, reader to readers.
Published:
一条社区转发把 Benchmark Radar 带往 X 更远处，读者传读者。
Published:
A Moltbook post carried Benchmark Radar onto the agent internet’s front page — the social feed built for agents, by agents.
Published:
一条 Moltbook 帖子把 Benchmark Radar 送上 agent 互联网的首页——那个为 agent 建、由 agent 用的社交 feed。
Published:
Alexy Khrabrov, who runs Strongly Typed AI News, shared its Benchmark Radar lead story on LinkedIn from his own account, carrying it from the news site into professional networks.
Published:
Strongly Typed AI News 的运营者 Alexy Khrabrov 用个人账号在 LinkedIn 转发了 Benchmark Radar 头条，把它从新闻站送进职业网络。
Published:
A launch-coverage share on LinkedIn carried the Benchmark Radar launch story into professional networks.
Published:
一条 LinkedIn 发布报道转发把 Benchmark Radar 发布故事送进职业网络。
Published:
The Isfahan AI Desk (میز هوش مصنوعی استان اصفهان), the province’s AI governance and ecosystem watch, listed our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation (arXiv:2609.11115) in its news radar with a relevance score of 91/100.
Published:
伊斯法罕省 AI 治理与生态观察（میز هوش مصنوعی استان اصفهان）把我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation（arXiv:2609.11115）列入新闻雷达，相关度评分 91/100。
Published:
Strongly Typed AI News, the AI news site Alexy Khrabrov runs at collected.ga, made our paper the lead story of its September 12, 2026 “Typesafe AI Daily” issue. Its crawler checked 49 sources that day and found 12 strong signals across AI infrastructure, funding, research, and developer tools, and it named Benchmark Radar the strongest of them.
Published:
Alexy Khrabrov 在 collected.ga 运营的 AI 新闻站 Strongly Typed AI News 把我们的论文做成 2026 年 9 月 12 日「Typesafe AI Daily」一期的头条。当天爬虫检查 49 个来源，在 AI 基建、融资、研究、开发者工具中找出 12 条强信号，Benchmark Radar 排第一。
Published:
Papers with Code listed our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation (arXiv:2609.11115), with its TL;DR, abstract, code repository, and project page.
Published:
Papers with Code 收录了我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation（arXiv:2609.11115），附 TL;DR、摘要、代码仓库与项目页。
Published:
Moonlight, the AI research-reading companion by Corca, published a literature-review page for our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation.
Published:
AI 论文阅读助手 Moonlight（Corca 出品）为我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 发布了文献解读页。
Published:
Geepity turned our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation into a full interactive reading edition, walking through all nine chapters with live charts and a working BM25F demo.
Published:
Geepity 把我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 做成完整互动解读版，九章走完，配实时图表与可运行的 BM25F 演示。
Published:
DailyAIWire covered the Benchmark Radar launch as a centralized AI evaluation database, with an AI-assisted intelligence report, an explain-like-I’m-five summary, and a visual flow of the system.
Published:
DailyAIWire 以中心化 AI 评测数据库为题报道了 Benchmark Radar 发布，附 AI 辅助情报分析、五岁小孩版解释与系统流程图。
Published:
arXivDaily, the Chinese arXiv digest with per-university pages, lists our paper on its Shanghai Jiao Tong University page, following the SJTU affiliation on the author list.
Published:
中文 arXiv 速递 arXivDaily 在上海交通大学高校专区收录了我们的论文，作者名单里的上交 affiliation 把它带到这里。
Published:
alphaXiv hosts a discussion page for our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation, with the abstract, author list, BibTeX citation, and open discussion.
Published:
alphaXiv 为我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation 开设了讨论页，含摘要、作者名单、BibTeX 引用与开放讨论区。
Published:
Papers with Code listed our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation (arXiv:2609.11115), with its TL;DR, abstract, code repository, and project page.
Published:
Papers with Code 收录了我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation（arXiv:2609.11115），附 TL;DR、摘要、代码仓库与项目页。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
One hundred twenty-six thousand eight hundred and eighty-seven dollars is housing-level money in Austin; in UT Austin’s new business-school building, it buys a single parking spot.
Published:
十二万六千八百八十七美元，在 Austin 已经是住房级别的钱；在 UT Austin 的新商学院大楼里，它只对应一个停车位。
Published:
In a field where everyone gets pulled toward the next hot thing, Juan Santiago joined Stanford Mechanical Engineering in 1998 and has been there ever since, building one of the most consequential microfluidics labs on the planet. Long-term focus is its own competitive advantage.
Published:
在一个人人都被拽向下一个热门方向的领域里，Juan Santiago 1998 年加入了斯坦福机械工程系，此后一直待在那里，建起了这个星球上最有影响力的微流控实验室之一。长期专注本身就是一种竞争优势。
Published:
A receipt in Singapore showed S$5.50, while a U.S. credit-card account displayed only US$4.27. Another purchase of about S$48 appeared as roughly US$38. A bank-card transit ride also failed to appear immediately as pending.
Published:
在新加坡消费时，一张收据写着 S$5.50，但美国信用卡账户只显示 US$4.27。另一笔约 S$48 的消费，则显示为约 US$38。与此同时，刷银行卡乘坐公共交通后，交易也没有立刻出现在 pending 中。
Published:
The last mile of an AI-generated travel plan may be one missing Retry button.
Published:
I was just thinking about why I can be polite to my brokerage’s customer service, but should not be polite to a middleman.
Published:
On February 1 this year, I wrote down a thought about one question: what kind of company is the greatest company in the world?
Published:
今年 2 月 1 日，我写过一段话，想的是一个问题，什么样的公司是世界上最伟大的公司。
Published:
I want the people close to me to live well. It is easy to let that wish turn into deciding what living well should mean for them.
Published:
我希望亲近的人过得好，也很容易顺着这个愿望，替他们想好怎样才算过得好。
Published:
A will is part of a love letter.
Published:
遗书其实是情书的一部分。
Published:
Soy milk is coffee. Youtiao is a croissant.
Published:
咖啡也是豆浆，牛奶也是豆浆。
Published:
I only recently found some of the feeling of being a leader again.
Published:
我最近才找回一些 leader 的感觉。
Published:
I don’t call a person brave lightly. The word is heavy, and close to praise. To say it is to affirm how someone faced fear, cost, and responsibility.
Published:
我不会随便说一个人勇敢。这个词太重，也太接近赞美；说出口，便是在肯定一个人如何面对恐惧、代价与责任。
Published:
O. Henry Museum, Austin
Published:
于奥斯汀，欧·亨利博物馆
Published:
A handcrafted website universe where every webpage, balloon, click, and line of code says “I love you” - designed exclusively for Momo. Visit Live Site
Published:
这是一个手工打造的网站小宇宙：每一个网页、气球、点击和每一行代码都在说“我爱你”——只为 Momo 而设计。访问在线站点
Published:
For a long time I thought adding thousands of people on LinkedIn by hand, just for fun, was all show and no use.
Published:
A few days ago Yann LeCun announced his new role in a LinkedIn post, something almost unthinkable in China.
Published:
There is no top-tier, universally applicable methodology, except one. Go deep in one industry, build personal influence, build a personal brand.
Published:
For technical people, a strong pull request on a fast-growing open-source repo can become better inbound marketing than another generic post about AI.
Published:
对技术人来说，在一个高速增长的开源仓库里做出高质量 PR，往往比再发一篇泛泛而谈的 AI 观点帖更像真正有效的 inbound marketing。
Published:
I spent a whole evening with Gemini taking apart the Buy Borrow Die playbook that American billionaires run, expecting that with a little scaling down I could just copy the moves, and what I found instead was that almost every single move has an F-1 trapdoor underneath it, and the list of things I can actually do fits on one page.
Published:
What I actually want to do today is pull myself out from under a mountain of repetitive work.
Published:
我今天真正想做的事情，其实是把自己从一座 mountains of 重复性的 work 里面捞出来。
Published:
A PhD is hard. A PhD in 2026 — with the field moving faster than your committee can read — is a different kind of hard. AI PhD Survival Guide is a handbook for surviving and finishing an AI or ML PhD without burning out and without falling behind.
Published:
读博很难。2026 年读博——当领域进展快到你的委员会都读不过来时——是另一种难。AI PhD Survival Guide 是一本手册，教你如何在 AI 或机器学习博士研究中活下来并毕业，既不会燃烧殆尽，也不至于掉队。
Published:
In the fall of 2015, two things happened simultaneously.
Published:
2015 年秋天，有两件事同时发生了。
Published:
Original Publication Date: March 15, 2026
Published:
Original Publication Date: March 15, 2026
Published:
A PhD system built on a 19th-century apprenticeship model is failing by nearly every empirical measure — ~40–50% attrition, depression rates six times the general population, and tenure-track placement below 15% in many fields — while venture capital has spent four decades perfecting bilateral contracts that manage exactly the risks PhD programs ignore: information asymmetry, moral hazard, hold-up, and misaligned incentives. The literature across economics, education policy, signaling theory, and AI research converges on a striking conclusion: the structural tools to fix the PhD already exist in VC contract design, but academia has never imported them. This research compendium maps the evidentiary landscape across six domains to ground the argument.
Published:
A year and a half into my PhD in the US, the best training I have received is learning to tell a good story.
Published:
A bit of history from a casual chat: in the US, science students used to get bigger stipends than business, law or medicine students, because science students were poor.
Published:
A student came to me today for advice on summer research applications, and when I asked what he wanted out of it, he said it straight: a return offer.
Published:
For kids who grew up in China, the first problem in applying abroad, and often the biggest, is not what to do. It is how they think.
Published:
NASA’s Future Investigators in NASA Earth and Space Science and Technology (FINESST) is one of the most underused fellowships among US graduate students. Most PhD students have never heard of it, and the ones who have often miss the deadline because the proposal expectations aren’t obvious from the call alone. This repo is a curated guide of links, tips, and examples to help you write a competitive FINESST proposal.
Published:
NASA 的未来地球与空间科技研究者（FINESST）项目，是美国研究生中最被低估的奖学金之一。这个仓库是一份精心整理的链接、技巧与示例指南，帮助你写出一份有竞争力的 FINESST 提案。
Published:
Whether a civilization has logic is not decided by whether it produced a few brilliant arguers. Three tests decide it. Did anyone study inference itself and write books about it. Did the subject enter schools, with teachers who taught it and students who learned it. Did some institution, an examination above all, force generation after generation to master it. Greece passes all three tests. India passes all three. Ancient China fails every one.
Published:
一个文明算不算有逻辑，不看它出过几个会推理的天才，看三样东西。有没有人把推理本身当成研究对象，写出专门的书。这门学问进没进学校，有没有老师教，学生学。有没有一套制度逼着一代代人学它，用它，考试就是最硬的制度。拿这三样去量希腊，量印度，都量得过。拿来量古代中国，三样全空。
Published:
I don’t call a person brave lightly. The word is heavy, and close to praise. To say it is to affirm how someone faced fear, cost, and responsibility.
Published:
我不会随便说一个人勇敢。这个词太重，也太接近赞美；说出口，便是在肯定一个人如何面对恐惧、代价与责任。
Published:
The promise of AI-automated science is intoxicating: imagine machines that can generate hypotheses, design experiments, and publish papers while we sleep.
Published:
AI 自动化科研的愿景令人陶醉：想象机器能在我们睡觉时生成假设、设计实验、发表论文。然而，尽管关于“AI 科学家”的新闻铺天盖地，我们正在撞上一堵根本性的墙。问题不在于算力或数据集规模，而是某种更深刻的东西，根植于科学发现的本质和信息论之中。
Published:
In October 2023 I asked ChatGPT an over-engineered question: how do you fry rice so that every grain of rice ends up bonded to egg, with no bare rice grains and no isolated clumps of egg sitting off on their own? I saved that conversation as an HTML file, dropped it in my Downloads folder, and did not look at it again for almost three years. Last week I finally turned it into an actual open-source repo, and going back through the original conversation to build it was more interesting than I expected.
Published:
2023 年 10 月，我问了 ChatGPT 一个过度工程化的问题：怎么炒蛋炒饭，才能让每一粒米都裹上蛋，既没有裸露的白米粒，也没有单独抱团、没沾到米的蛋碎？那次对话我存成了一个 HTML 文件，扔进 Downloads 文件夹，然后差不多三年没再打开过。上周我终于把它整理成了一个真正的开源仓库，回头重读那段 2023 年的对话、把它做成代码的过程，比我预想的有意思得多。
Published:
Top physics researchers aren’t blind to what AI is doing in the application layer. They’re locked by their own evaluation system and can’t afford to look.
Published:
How photographing a black hole and training GPT-4 are fundamentally the same process: extracting coherent structure from impossibly sparse frequency-domain measurements.
Published:
你好，我是 Koutian。第十五天，雷达第一次学会主动对外说话，不再只是等你来翻。
Published:
You can study fluid mechanics in five different countries before you turn 30, work on petroleum reservoirs and tectonophysics and planetary ice on the same week, and somehow end up at a Centennial Chair in Geophysics. Marc Hesse did exactly that, and the through-line is more interesting than any of the individual stops.
Published:
你可以在 30 岁前在五个不同国家学流体力学，同一周里既研究油气储层又研究构造物理和行星冰，最后竟然坐上地球物理学百年讲席。Marc Hesse 就是这么做的，而他背后那条主线比任何一个单独的站点都更有意思。
Published:
NASA’s Future Investigators in NASA Earth and Space Science and Technology (FINESST) is one of the most underused fellowships among US graduate students. Most PhD students have never heard of it, and the ones who have often miss the deadline because the proposal expectations aren’t obvious from the call alone. This repo is a curated guide of links, tips, and examples to help you write a competitive FINESST proposal.
Published:
NASA 的未来地球与空间科技研究者（FINESST）项目，是美国研究生中最被低估的奖学金之一。这个仓库是一份精心整理的链接、技巧与示例指南，帮助你写出一份有竞争力的 FINESST 提案。
Published:
On May 17, 2025, I had about nine hours alone in Hong Kong between flights and trains, and I planned them around Victoria Harbour and the city’s richest commercial districts.
Published:
2025年5月17日，我一人途经香港，可游览时间约九个小时，目标是维多利亚港和香港最繁华的商业地段。
Published:
Most land surface models treat a tree like a passive straw. Water comes in at the roots, water leaves at the leaves, end of story. Ashley Matheny’s research basically says no, a tree is an active hydraulic system with storage, capacitance, and a strategy, and if you do not model it that way you are going to be wrong about drought.
Published:
大多数陆地表面模型都把一棵树当成一根被动的吸管。水从根进来，水从叶出去，故事就这么简单。Ashley Matheny 的研究基本上在说：不对，树是一个带有储水、电容和策略的活跃水力系统，如果你不这样建模，你在干旱问题上就会犯错误。
Published:
If you ask most people where trees in California get their water in a drought, they will say “the soil.” It turns out a huge fraction of it comes from cracks in the bedrock underneath the soil, and Daniella Rempe is the person who put numbers on it.
Published:
如果你问大多数人，加州在干旱时树从哪里取水，他们会说”土壤”。事实证明，很大一部分水其实来自土壤下方岩石裂缝里的基岩，而 Daniella Rempe 就是把数字放到这件事上的人。
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
Published:
Jiayi Weng planned his career at OpenAI around one goal, to get his name onto as many OpenAI blog posts as possible.
Published:
O. Henry Museum, Austin
Published:
于奥斯汀，欧·亨利博物馆
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
Most academic climate scientists will tell you they care about policy and then publish a paper that no policymaker is ever going to read. Geeta Persad spent four years working at the Union of Concerned Scientists translating climate models for water managers, and she came back to academia knowing exactly what the gap looks like.
Published:
大多数学术气候科学家会告诉你他们在乎政策，然后发表一篇任何决策者都不会读的论文。Geeta Persad 在忧思科学家联盟（Union of Concerned Scientists）花了四年，为水资源管理者翻译气候模型，然后带着对那个差距的清醒认识回到了学术界。
Published:
Why AI agents lack temporal awareness, and how injecting timestamps transforms them from stateless automata into coherent entities.
Published:
The sharpest minds keep moving into AI labs, and it isn’t just for the money.
Published:
这是《颠覆性创新者的思维模式》的五个月后续研究。
Published:
I just remembered something sudden on the train: why are almost all the links on the internet that extremely saturated shade of blue?
Published:
我刚刚在火车上突然想起一件事：为什么互联网上的链接，几乎清一色都是那种饱和度极高的蓝色？
Published:
Top physics researchers aren’t blind to what AI is doing in the application layer. They’re locked by their own evaluation system and can’t afford to look.
Published:
When gold gets too expensive relative to silver, you sell your gold and buy silver. When silver catches up, you swap back. This sounds like folk wisdom, but it’s one of the most time-tested strategies in precious metals.
Published:
建立在 19 世纪学徒制模型上的博士培养体系，几乎在所有关键经验指标上都已经出现了系统性失灵。
Published:
Original Publication Date: March 15, 2026
Published:
Original Publication Date: March 15, 2026
Published:
A PhD system built on a 19th-century apprenticeship model is failing by nearly every empirical measure — ~40–50% attrition, depression rates six times the general population, and tenure-track placement below 15% in many fields — while venture capital has spent four decades perfecting bilateral contracts that manage exactly the risks PhD programs ignore: information asymmetry, moral hazard, hold-up, and misaligned incentives. The literature across economics, education policy, signaling theory, and AI research converges on a striking conclusion: the structural tools to fix the PhD already exist in VC contract design, but academia has never imported them. This research compendium maps the evidentiary landscape across six domains to ground the argument.
Published:
Parameterizing a large Fortran climate model by hand is slow, error-prone, and hard to validate. Noah-Agent asks whether a team of specialized AI agents can do it instead.
Published:
手动为大型 Fortran 气候模型配置参数，速度慢、容易出错，也难以验证。Noah-Agent 要检验的是，一组各有所长的 AI 智能体能否协作完成这项工作。
Published:
A year and a half into my PhD in the US, the best training I have received is learning to tell a good story.
Published:
The Isfahan AI Desk (میز هوش مصنوعی استان اصفهان), the province’s AI governance and ecosystem watch, listed our paper Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation (arXiv:2609.11115) in its news radar with a relevance score of 91/100.
Published:
伊斯法罕省 AI 治理与生态观察（میز هوش مصنوعی استان اصفهان）把我们的论文 Benchmark Radar: A Living Database and Search Engine for AI Benchmarks and Evaluation（arXiv:2609.11115）列入新闻雷达，相关度评分 91/100。
Published:
DailyAIWire covered the Benchmark Radar launch as a centralized AI evaluation database, with an AI-assisted intelligence report, an explain-like-I’m-five summary, and a visual flow of the system.
Published:
DailyAIWire 以中心化 AI 评测数据库为题报道了 Benchmark Radar 发布，附 AI 辅助情报分析、五岁小孩版解释与系统流程图。
Published:
Secretly recording a customer interview can turn a useful research habit into a privacy problem in one click.
Published:
偷偷录一次客户访谈，可能把一个正常的产品研究流程变成隐私和合规事故。
Published:
Exchanging money for time is a privilege I rarely indulge in, but a surprise membership benefit that cleared airport security in ten minutes changed my perspective on friction and market regulation. While a Clear Plus membership normally costs over a hundred dollars annually, obtaining it for free through an Uber membership allowed me to experience a level of efficiency that money can’t always buy—at least not without a well-regulated system behind it.
Published:
用金钱换取时间是一种我很少尝试的“特权”，但这次在奥斯汀机场仅用十分钟便完成安检的经历，让我对效率与市场调节有了新的思考。这份特权并非我主动购买，而是通过 Uber 年费会员赠送的 Clear Plus 获得的——原本需要每年支付一百多美金的服务，在免除门槛后，带给我一种金钱也未必能随时买到的流畅体验。
Published:
Seizing the gateway to the human mind in the AI era is, at bottom, a secret war over token throughput power.
Published:
抢占 AI 时代的人类心智入口，说到底是一场关于 token 吞吐权力的隐秘战争。
Published:
Lay out every hyped multiAgent product on a table in front of Steve Jobs and he would probably squint, frown, and ask one question, why on earth are you showing the user any of this.
Published:
把现在市面上所有被吹得天花乱坠的multiAgent产品摆在乔布斯面前，他大概率会皱着眉头反问一句，你为什么要让用户看到这些东西。
Published:
Five YC rebels: some were expelled, some were welcomed back. What happened next?
Published:
This is the worst startup product I have seen.
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
Published:
Yuxuan and I have been discussing AI agent ideas for a while. Yesterday, we finally decided: we’re building Atomize, a task-breaking agent system evolved from Goblin Tools.
Published:
我和 Yuxuan 讨论 AI Agent 的想法已经有一段时间了。昨天，我们终于做出了决定：我们要做 Atomize，一个从 Goblin Tools 演化而来的任务分解 Agent 系统。
Published:
This is the worst startup product I have seen.
Published:
你好，我是 Koutian。第十四天，我们把雷达的几个老毛病修好了，也让它更像一个会自己干活的工具。
Published:
Goal: Use a 4x3 grid to place every relevant person “by position, with a matching strategy,” then review it regularly to maximize resource use and manage risk.
Published:
To use OpenClaw or not to use OpenClaw, that is the question.
Published:
用不用 OpenClaw？这是个问题
Published:
Every knowledge worker is now competing with an LLM for some part of their job. AI Survival Guide is a handbook for figuring out which parts, what to do about it, and how to come out the other side better at your craft instead of replaced by it.
Published:
如今，每一位知识工作者都在用某项工作的某些部分与 LLM 竞争。AI Survival Guide 是一本手册，帮你弄清是哪些部分、该怎么应对，以及如何带着更强的技艺从这场竞争的另一头走出来，而不是被它取代。
Published:
Every generation has its own way of managing work: paper notebooks, SaaS task managers, and now programmable agentic workflows powered by tools like OpenClaw heartbeat.
Published:
每一代人都有自己管理任务的方式：最早是纸和笔，后来是 SaaS 任务管理工具，现在则开始进入像 OpenClaw heartbeat 这样可程序化、可持续运行的 agentic workflow 时代。
Published:
Early-career growth usually comes down to a few repeatable layers: alignment, delivery, visibility, communication, leadership, and trust.
Published:
职业早期的成长，通常可以拆成几个可重复的层次：方向对齐、稳定交付、建立可见度、提升沟通、展现领导力，以及赢得信任。
Published:
The most important measure of work may be the energy it consumes, more than the time it takes.
Published:
衡量工作效率，最重要的指标或许不是投入的时间，而是消耗的能源。
Published:
Yuxuan and I have been discussing AI agent ideas for a while. Yesterday, we finally decided: we’re building Atomize, a task-breaking agent system evolved from Goblin Tools.
Published:
我和 Yuxuan 讨论 AI Agent 的想法已经有一段时间了。昨天，我们终于做出了决定：我们要做 Atomize，一个从 Goblin Tools 演化而来的任务分解 Agent 系统。
Published:
Don’t expect your first video to go viral, and don’t expect your first sales trip to land ten million dollars, but you can ask what happens after the hundredth try.
Published:
“这b人生过的值不值” — roughly, “was this damn life worth it?” — is a phrase that does a lot of emotional work on the Chinese internet. ZuoGuoHuaDiao (做过划掉) is a small web tool that takes the phrase seriously and tries to answer it with numbers. Try Live Demo
Published:
“这b人生过的值不值”——简单说，就是”这该死的一生过得值不值？”——这句在中国互联网上承载了大量情感的话语。ZuoGuoHuaDiao（做过划掉）是一款小巧的网页工具，它认真对待这句话，并试图用数字来回答它。试试在线演示
Published:
Most people think of summer vacation as time off. I started wondering whether that framing undersells it. Summer Calculator is a small web tool that estimates the full value of your summer — money, learning, relationships, health — not just the days you didn’t work. Try Live Demo
Published:
大多数人把暑假当作休息时间。我开始怀疑这种框架是否低估了它。Summer Calculator 是一款小巧的网页工具，用来估算你暑假的完整价值——金钱、学习、人际关系、健康——而不仅仅是你没工作的那些日子。试试在线演示
Published:
Muck Rack now has a profile for me. Muck Rack is a media database that journalists and communications teams use to find people and track who has written or been written about.
Published:
Muck Rack 现在有了我的个人页面。Muck Rack 是一个媒体数据库，记者和公关团队用它找人，也用它查谁写过什么、谁被写过。
Published:
HyperAI now lists my work. HyperAI is an AI community site that collects papers, datasets, and tutorials, with an English and a Chinese edition.
Published:
HyperAI 超神经 现在收录了我的工作。HyperAI 是一个 AI 社区网站，汇集论文、数据集和教程，有中英文两个版本。
Published:
Shenzhen X-Institute interviewed me for 寻找X-学子说：吴叩天：从零到一、创造无限. The conversation covers how I found research, learned to build projects, and thought about creating work from zero to one.
Published:
深圳零一学院在《寻找X-学子说：吴叩天：从零到一、创造无限》中采访了我。对话谈到我如何找到科研方向、学习把想法做成项目，也谈到对“从零到一”的理解。
Published:
My Semantic Scholar profile indexes my research and connects it with citations, references, and related work.
Published:
我的 Semantic Scholar 主页收录了研究成果，也把论文与引用、参考文献和相关工作连接起来。
Published:
My Scopus author profile indexes my publications and citation information in the Scopus database.
Published:
我的 Scopus 作者主页收录了论文及其在 Scopus 数据库中的引用信息。
Published:
My ResearchGate profile gathers my publications on a research-focused network where readers can follow projects and authors.
Published:
我的 ResearchGate 主页汇总了论文，读者也可以在这个科研社交平台上关注项目和作者。
Published:
My ORCID record provides a persistent researcher identifier and connects my work across publishers and research systems.
Published:
我的 ORCID 记录提供长期有效的研究者标识，并把不同出版社和科研系统中的成果关联起来。
Published:
My Google Scholar profile collects my publications and citation record. It is the quickest place to see new papers and follow how they are cited.
Published:
我的 Google Scholar 主页汇总了论文和引用记录。想查看最新论文及其引用情况，可以从这里开始。
Published:
CatalyzeX now has an author page for me. CatalyzeX pairs machine learning papers with the code that goes with them, so a reader who finds a paper can jump straight to a working version instead of rebuilding it from the PDF.
Published:
CatalyzeX 现在有了我的作者页面。CatalyzeX 把机器学习论文和对应的代码配在一起，读者找到论文后可以直接跳到能跑的版本，不用照着 PDF 从头重写。
Published:
AIModels.fyi now has an author page for me. The site turns new AI research papers into short summaries in plain English, so people who build with AI can keep up without reading every PDF.
Published:
AIModels.fyi 现在有了我的作者页面。这个网站把新发表的 AI 论文写成简短的大白话摘要，让用 AI 做产品的人不必逐篇读 PDF 也能跟上进展。
Published:
500AI keeps an index of AI researchers, and it now has a page for me. It lists me as a Ph.D. student in Geological Sciences and a graduate research assistant at the University of Texas at Austin, tags my work under information retrieval and evaluation, and links out to this site, GitHub, LinkedIn, X, Google Scholar, and my UT Jackson School page.
Published:
500AI 是一个 AI 研究者索引，现在收录了我的页面。页面写明我是得克萨斯大学奥斯汀分校地质科学博士生、研究助理，把我的方向标为信息检索与评测，并链接到本站、GitHub、LinkedIn、X、Google Scholar 和 UT Jackson School 主页。
Published:
I use X for short research notes, project updates, and links to new work.
Published:
My UT Jackson School profile introduces my doctoral work in the Department of Earth and Planetary Sciences at The University of Texas at Austin.
Published:
我的 UT Jackson School 主页介绍了我在得克萨斯大学奥斯汀分校地球与行星科学系的博士研究。
Published:
I contribute to Science, Y’all!, the Jackson School’s platform for making earth and planetary science accessible to broader audiences.
Published:
我参与 Science, Y’all! 的科学传播工作。这个 Jackson School 平台面向更广泛的读者介绍地球与行星科学。
Published:
My publications index gathers my papers across earth science, AI for science, and AI evaluation. Each entry links to the work and its available project resources.
Published:
My LinkedIn profile is where I share research, projects, and professional updates.
Published:
我在 LinkedIn 分享研究、项目和职业动态。
Published:
My Instagram profile collects visual notes from research, travel, projects, and daily life.
Published:
我的 Instagram 主页记录科研、旅行、项目和日常生活中的图像片段。
Published:
My GitHub profile is the main index of my open-source work, research code, websites, and experiments.
Published:
我的 GitHub 主页集中展示开源项目、研究代码、网站和各种实验。
Published:
Tacite is one of my projects exploring AI products and how people interact with them.
Published:
ResearchSkills.ai is a directory for reusable AI skills built for research workflows. It helps researchers find and share tools for literature work, analysis, writing, and other parts of the research process.
Published:
ResearchSkills.ai 收录可以复用的科研 AI Skills，方便研究者查找和分享用于文献、分析、写作等研究环节的工具。
Published:
ResearchClawBench evaluates whether AI agents can carry out end-to-end scientific research tasks.
Published:
ResearchClawBench 评测 AI 智能体能否完成端到端的科学研究任务。
Published:
My personal site brings together my research, projects, writing, and media links in one place.
Published:
MatrAIx is a population-scale simulated-user evaluation framework for testing AI systems and digital products with diverse persona agents.
Published:
Earth-Space-AI is an open-source organization for AI tools, skill packages, and research projects in earth and space science.
Published:
Earth-Space-AI 是一个面向地球与空间科学的开源组织，收录 AI 工具、Skill 包和研究项目。
Published:
ASI-Bench evaluates AI agents on 60 research-grade tasks across 11 scientific domains.
Published:
My academic website is the central index for my research, publications, projects, writing, and media coverage.
Published:
Crypto Dashboard is a browser-based dashboard for following cryptocurrency market data and trends in one place.
Published:
Crypto Dashboard 是一个在浏览器中集中查看加密货币市场数据和趋势的仪表盘。
Published:
An AI writing assistant that polishes academic text using DeepSeek-R1 and detects AI-generated content.
Published:
学术写作的严谨性不仅在于数据，更在于表达的精准，而 1AI-polish 通过集成 DeepSeek-R1 的推理能力，为研究者提供了一套集文本润色与 AI 检测于一体的深度协作系统，旨在让复杂的科研思想以更专业、更透明的方式呈现。
Published:
UT01 is a free, student-built collection of resources for The University of Texas at Austin. It brings practical campus information and community projects into one public site.
Published:
Prompt Caching Is Not Technical Debt. It’s what the math demands.
Published:
很多人把 prompt caching 看成一个省钱 hack，但我觉得这个判断刚好反了。
Published:
The most important measure of work may be the energy it consumes, more than the time it takes.
Published:
衡量工作效率，最重要的指标或许不是投入的时间，而是消耗的能源。
Published:
NASA’s Future Investigators in NASA Earth and Space Science and Technology (FINESST) is one of the most underused fellowships among US graduate students. Most PhD students have never heard of it, and the ones who have often miss the deadline because the proposal expectations aren’t obvious from the call alone. This repo is a curated guide of links, tips, and examples to help you write a competitive FINESST proposal.
Published:
NASA 的未来地球与空间科技研究者（FINESST）项目，是美国研究生中最被低估的奖学金之一。这个仓库是一份精心整理的链接、技巧与示例指南，帮助你写出一份有竞争力的 FINESST 提案。
Published:
Scientists sell themselves too.
Published:
科学家也是出来卖的。
Published:
One hundred twenty-six thousand eight hundred and eighty-seven dollars is housing-level money in Austin; in UT Austin’s new business-school building, it buys a single parking spot.
Published:
十二万六千八百八十七美元，在 Austin 已经是住房级别的钱；在 UT Austin 的新商学院大楼里，它只对应一个停车位。
Published:
I am deeply honored to join my colleagues in contributing to the newly published paper, “On the Ethics of Generative GeoAI: Explainability, Bias, Hallucination, Accountability, Privacy, and Trust,” specifically focusing on the section regarding GeoAI Trust (Section 8).
Published:
我非常荣幸能够参与到这篇新发表的论文 《On the Ethics of Generative GeoAI: Explainability, Bias, Hallucination, Accountability, Privacy, and Trust》 的合作中，并作为共同作者之一，贡献了关于 GeoAI信任（GeoAI Trust） 的章节（第八章）。
Published:
If you’re serious about understanding LLMs and AI engineering, Simon Willison’s Newsletter is one of the best resources out there. With over 38,000 subscribers, it provides detailed analysis on AI, LLMs, web engineering, open source, data science, and Python.
Published:
如果你认真想要理解 LLM 和 AI 工程，Simon Willison 的 Newsletter 是最好的资源之一。拥有超过 38,000 名订阅者，它提供关于 AI、LLM、Web 工程、开源、数据科学和 Python 的深度分析。
Published:
TQQQ is the 3x-leveraged Nasdaq-100 ETF. It is also one of the most asymmetric instruments retail investors touch — the upside is real, the drawdowns are brutal, and the daily-rebalance math means buy-and-hold doesn’t behave the way most people assume. TQQQ ML Trend is an experiment in using machine learning to predict the trend regime, not the price.
Published:
TQQQ 是纳斯达克 100 指数的 3 倍杠杆 ETF。它也是散户接触过的最不对称的金融工具之一——上行空间真实存在，回撤却十分残酷，而每日再平衡的数学逻辑意味着”买入并持有”并不像大多数人以为的那样运转。TQQQ ML Trend 是一个用机器学习来预测趋势状态的实验，而不是去预测价格。
Published:
Python and MATLAB tools for creating publication-quality scientific maps and geographic visualizations.
Published:
用 Python 和 MATLAB 创建适合论文发表的科学地图与地理可视化图件。
Published:
A web-based platform for facilitating cross-cultural communication and understanding.
Published:
一个基于 Web 的平台，用于促进跨文化交流与理解。
Published:
第十三天的重点是让仪表盘更好看，并且把每条结论都说得有凭有据。
Published:
Another day of heavy internet surfing, and the biggest lesson was this: there is no investment strategy that lets you lie down and get financially free.
Published:
TQQQ is the 3x-leveraged Nasdaq-100 ETF. It is also one of the most asymmetric instruments retail investors touch — the upside is real, the drawdowns are brutal, and the daily-rebalance math means buy-and-hold doesn’t behave the way most people assume. TQQQ ML Trend is an experiment in using machine learning to predict the trend regime, not the price.
Published:
TQQQ 是纳斯达克 100 指数的 3 倍杠杆 ETF。它也是散户接触过的最不对称的金融工具之一——上行空间真实存在，回撤却十分残酷，而每日再平衡的数学逻辑意味着”买入并持有”并不像大多数人以为的那样运转。TQQQ ML Trend 是一个用机器学习来预测趋势状态的实验，而不是去预测价格。
Published:
A hand-controlled 3D visualization of your AI conversation history - fly through ChatGPT conversations and explore artificial intelligence concepts. Try Live Demo
Published:
一个支持手势控制的 3D 可视化项目，把你的 AI 对话历史变成可以飞行探索的星系。你可以在 ChatGPT 对话之间穿梭，也可以把抽象的人工智能概念变成一个能亲手操作的可视化空间。Try Live Demo
Published:
Most crypto dashboards either drown you in numbers or hide the signal behind a paywall. Crypto Dashboard is my attempt at a clean, AI-augmented view of the market that runs entirely in your browser. Try Live Demo
Published:
大多数加密货币仪表盘要么把你淹没在数字里，要么把信号藏在付费墙后面。Crypto Dashboard 是我对一种简洁、由 AI 增强的市场视图的尝试，而且它完全运行在你的浏览器里。试试在线演示
Published:
A comprehensive tool built with React, Vite, and Supabase to help individuals plan and navigate their careers in the AI industry.
Published:
这是一个基于 React、Vite 和 Supabase 构建的综合工具，帮助个人规划并导航自己在 AI 行业中的职业发展路径。
Published:
On February 1 this year, I wrote down a thought about one question: what kind of company is the greatest company in the world?
Published:
今年 2 月 1 日，我写过一段话，想的是一个问题，什么样的公司是世界上最伟大的公司。
Published:
I want the people close to me to live well. It is easy to let that wish turn into deciding what living well should mean for them.
Published:
我希望亲近的人过得好，也很容易顺着这个愿望，替他们想好怎样才算过得好。
Published:
A will is part of a love letter.
Published:
遗书其实是情书的一部分。
Published:
Soy milk is coffee. Youtiao is a croissant.
Published:
咖啡也是豆浆，牛奶也是豆浆。
Published:
I only recently found some of the feeling of being a leader again.
Published:
我最近才找回一些 leader 的感觉。
Published:
I don’t call a person brave lightly. The word is heavy, and close to praise. To say it is to affirm how someone faced fear, cost, and responsibility.
Published:
我不会随便说一个人勇敢。这个词太重，也太接近赞美；说出口，便是在肯定一个人如何面对恐惧、代价与责任。
Published:
O. Henry Museum, Austin
Published:
于奥斯汀，欧·亨利博物馆
Published:
Starting a petition, founding a club, and joining a research effort someone else organized look like three different things. They all run into the same question: who decides, who puts in the time, and when something goes wrong, who bears the consequences?
Published:
发起请愿、成立社团、参加别人组织的科研活动，看上去是三件不同的事。它们都绕不开同一个问题：谁作决定，谁投入时间，事情出了问题以后，谁承担后果？
Published:
A laser lightsail accelerating to one percent of light speed turns about one percent of the beam into motion. The other ninety-nine percent leaves as light and does no work.
Published:
激光光帆加速到光速的百分之一，光束里大约只有百分之一变成了运动。剩下的百分之九十九以光的形式离开，不做功。
Published:
Today, alumnus Pan Jianwei sent his spiritual genes (his name) beyond the Earth.
Published:
今天，校友潘建伟把自己的精神基因（名字）传播到了地球之外。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
If I reopened the 2026 folder in 2030, I would check one thing first: did the work completed with AI change a real problem and leave evidence that others could verify and build upon?
Published:
假设在 2030 年重新打开 2026 年的文件夹，我会先查一件事：今天借助 AI 完成的工作，是否改变了一个现实问题，并留下别人能够复验和继承的证据。
Published:
The thing that almost stranded me on this international trip was not a typhoon, but several unrelated systems deciding at the same time that they did not trust me.
Published:
这次国际旅行真正差点把我留在原地的，不是台风，而是几个互不认识的系统同时决定不相信我。
Published:
Going around GitHub opening issues and pull requests is sperm donation. The same mechanism, running as-is.
Published:
在 GitHub 上到处开 issue、提 pull request，就是赛博捐精。同一套机制，原样运行。
Published:
Let me offer a comparison that’s a bit inappropriate but does capture a real structure of open-source collaboration:
Published:
我提一个不太恰当、但确实抓住了开源协作某种结构的比喻：
Published:
I just remembered something sudden on the train: why are almost all the links on the internet that extremely saturated shade of blue?
Published:
我刚刚在火车上突然想起一件事：为什么互联网上的链接，几乎清一色都是那种饱和度极高的蓝色？
Published:
Two years later, I am back in this familiar city.
Published:
In my first post on this I described sending a patch through GitGitGadget and watching it land on the mailing list as v1. That post ended with the patch waiting for review. It has since merged. Now that I have seen the full lifecycle, from typo to master, here is what actually mattered.
Published:
在第一篇文章里，我写了怎么通过 GitGitGadget 把补丁发出去，它作为 v1 落在了邮件列表上。那篇文章结束在等待 review 的状态。现在它已经合并了。看完从一个拼写错误到进入 master 的完整过程，下面是我觉得真正重要的几点。
Published:
Recently I’ve sometimes felt confused by how far AI has come, a bit lost and anxious, unsure what I should do, and then, because of my identity and my path, wondering what fallbacks or better options exist.
Published:
最近虽然有的时候也会因为 AI 的发展程度感到非常困惑，感到有一点迷茫、有些焦虑，不知道自己该干啥，然后又因为自己的身份和路径问题，在想有什么样的退路或者更好的方案。
Published:
Today I got my first Git patch onto the mailing list.
Published:
今天我差一点，就把自己的第一个 Git 项目补丁送进邮件列表了。
Published:
While you’re still grinding your heart out on “if you just work hard enough, you can move up,” you may not realize the system was never designed to cultivate you in the first place, it was designed to screen you.
Published:
当你还在为“只要努力就能向上流动”而拼命内卷时，你可能没有意识到，这个系统从一开始就不是为了培养你，而是为了筛选你。
Published:
After two years of “All-in AI”, my deep-seated belief in the “10,000-hour talent moat” was ruthlessly shattered by the rise of Agentic workflows.
Published:
在“All-in AI”两年后，我曾经深信的“10,000小时人才护城河”理论，在Agentic工作流的冲击下被无情地打碎了一半。
Published:
Hah, I can’t help but laugh. It just hit me: the first time I ever used Markdown was back when I was building a poetry society. And that’s also when I first learned about Git. Looking back now, if you put all the founding members of that poetry society together, you’d almost have two unicorns.
Published:
哎，我他妈笑了。我忽然想起来，我最早用 Markdown，就是之前创建诗社的时候。知道 Git，也是在那个时候。现在回头看，整个诗社的元老凑在一起，真的快有两个独角兽了。
Published:
Earlier today I published a piece called “Four Ego Mistakes I Made as a 22-Year-Old Founder.” Then I ran it through Gemini 2.5 Pro as an independent reviewer. Gemini’s verdict was that the essay is itself an ego move. I think Gemini is mostly right.
Published:
今天早些时候我发了一篇叫《22 岁 founder 的四个 ego 错误》。然后我把这篇喂给了 Gemini 2.5 Pro 当独立审稿人。Gemini 的判决是：这篇文章本身就是一个 ego 动作。我觉得 Gemini 大体是对的。
Published:
If you want to feel your own growth directly, write a diary every day.
Published:
“这b人生过的值不值” — roughly, “was this damn life worth it?” — is a phrase that does a lot of emotional work on the Chinese internet. ZuoGuoHuaDiao (做过划掉) is a small web tool that takes the phrase seriously and tries to answer it with numbers. Try Live Demo
Published:
“这b人生过的值不值”——简单说，就是”这该死的一生过得值不值？”——这句在中国互联网上承载了大量情感的话语。ZuoGuoHuaDiao（做过划掉）是一款小巧的网页工具，它认真对待这句话，并试图用数字来回答它。试试在线演示
Published:
Jiayi Weng planned his career at OpenAI around one goal, to get his name onto as many OpenAI blog posts as possible.
Published:
Goal: Use a 4x3 grid to place every relevant person “by position, with a matching strategy,” then review it regularly to maximize resource use and manage risk.
Published:
I want the people close to me to live well. It is easy to let that wish turn into deciding what living well should mean for them.
Published:
我希望亲近的人过得好，也很容易顺着这个愿望，替他们想好怎样才算过得好。
Published:
Research date: 2026-05-26
Published:
研究时间：2026-05-26
Published:
你好，我是 Koutian。第十四天，我们把雷达的几个老毛病修好了，也让它更像一个会自己干活的工具。
Published:
Port Aransas South Jetty published an Institute Insights feature on May 28, 2026 about how our University of Texas team is using satellite remote sensing and AI to monitor Texas coastal water quality.
Published:
阿兰瑟斯港南防波堤 于 2026 年 5 月 28 日发表了一篇 Institute Insights 专题，讲的是我们德克萨斯大学团队如何用卫星遥感与 AI 监测德克萨斯沿海水质。
Published:
机器之心 (Synced) covered MatrAIx, the population-scale simulated-user evaluation framework I coauthored. The article introduces Persona 8B and the environments where persona agents can answer surveys, talk with chatbots, browse websites, and use apps.
Published:
机器之心报道了我参与的研究 MatrAIx。文章介绍了 Persona 8B，以及人格体智能体回答问卷、与聊天机器人对话、浏览网页和操作应用的四类测试环境。
Published:
机器之心 (Synced) covered ASI-Bench, a benchmark I coauthored for evaluating AI agents on scientific research. The benchmark contains 60 research-grade tasks across 11 scientific domains.
Published:
机器之心报道了我参与的科研智能体基准 ASI-Bench。这个基准收录了 11 个科学领域的 60 项研究级任务，用来检验 AI 智能体处理真实科研问题的能力。
Published:
Planetary Science featured our research on the spatial and temporal distribution of meteor speed and flux observed by a midlatitude meteor radar. The work uses radar observations to study how meteors vary across time and viewing geometry.
Published:
“行星科学”介绍了我们利用中纬度流星雷达研究流星速度和通量时空分布的工作。这项研究通过雷达观测，分析流星活动随时间和观测方向的变化。
Published:
WIRED Czech introduced MatrAIx and Persona 8B to its readers. The short News Beat item highlights 8.3 billion digital profiles described across 1,290 categorical dimensions.
Published:
WIRED Czech向读者介绍了 MatrAIx 和 Persona 8B。这篇短报道重点提到 83 亿条数字人格体记录，以及用于描述它们的 1,290 个类别维度。
Published:
NZZ am Sonntag featured MatrAIx in a story about eight billion digital counterparts. The coverage introduces the scale of Persona 8B and the idea of using diverse simulated users to evaluate products and AI systems.
Published:
NZZ am Sonntag在一篇关于“八十亿数字分身”的文章中介绍了 MatrAIx。报道从 Persona 8B 的规模出发，介绍了用不同背景的模拟用户测试产品和 AI 系统的思路。
Published:
Numerama published a detailed explanation of MatrAIx. It distinguishes the 8.3 billion persona records from simultaneously running agents and discusses model dependence, representativeness, privacy, and the role of human panels.
Published:
Numerama详细介绍了 MatrAIx。文章区分了 83 亿条人格体记录和同时运行的智能体，也讨论了模型依赖、代表性、隐私以及真人测试小组的作用。
Published:
Hugging Face Daily Papers featured MatrAIx as the number-two paper of the day. Its paper page collects the abstract, author list, discussion, and links to project resources in one place.
Published:
Hugging Face Daily Papers 收录了 MatrAIx，并将其列为当日第二名论文。论文页集中展示了摘要、作者名单、讨论和项目资源链接。
Published:
Forbes Türkiye examined how MatrAIx could help test health applications before they reach patients. The article focuses on variation in age, language, health literacy, and accessibility, while keeping clinical responsibility with people.
Published:
Forbes Türkiye讨论了 MatrAIx 如何在医疗应用接触真实患者前帮助测试。文章重点关注年龄、语言、健康素养和无障碍需求带来的差异，同时明确临床责任仍应由人承担。
Published:
Forbes covered MatrAIx, a study I coauthored, as a new research instrument for psychology. The article explains how population-scale AI personas can make experiments faster and easier to repeat while stressing that simulated users are not substitutes for people.
Published:
Forbes 报道了我参与的研究 MatrAIx，并讨论了 AI 人格体在心理学研究中的用途。文章指出，这类工具能让实验更快、更容易重复，但模拟用户不能代替真人。
Published:
Starting a petition, founding a club, and joining a research effort someone else organized look like three different things. They all run into the same question: who decides, who puts in the time, and when something goes wrong, who bears the consequences?
Published:
发起请愿、成立社团、参加别人组织的科研活动，看上去是三件不同的事。它们都绕不开同一个问题：谁作决定，谁投入时间，事情出了问题以后，谁承担后果？
Published:
I am deeply honored to join my colleagues in contributing to the newly published paper, “On the Ethics of Generative GeoAI: Explainability, Bias, Hallucination, Accountability, Privacy, and Trust,” specifically focusing on the section regarding GeoAI Trust (Section 8).
Published:
我非常荣幸能够参与到这篇新发表的论文 《On the Ethics of Generative GeoAI: Explainability, Bias, Hallucination, Accountability, Privacy, and Trust》 的合作中，并作为共同作者之一，贡献了关于 GeoAI信任（GeoAI Trust） 的章节（第八章）。
Published:
Top physics researchers aren’t blind to what AI is doing in the application layer. They’re locked by their own evaluation system and can’t afford to look.
Published:
Just by running a single command while using AI for research, you can turn your tacit knowledge into a Nature co-authorship.
Published:
只要你在用AI辅助科研，敲一行命令就能把你脑子里的「隐性经验」变成能发Nature的资本。
Published:
A PhD is hard. A PhD in 2026 — with the field moving faster than your committee can read — is a different kind of hard. AI PhD Survival Guide is a handbook for surviving and finishing an AI or ML PhD without burning out and without falling behind.
Published:
读博很难。2026 年读博——当领域进展快到你的委员会都读不过来时——是另一种难。AI PhD Survival Guide 是一本手册，教你如何在 AI 或机器学习博士研究中活下来并毕业，既不会燃烧殆尽，也不至于掉队。
Published:
Research, stock trading, and startups all share a common pattern: success comes from combining a defensible core competency with the ability to ride trending waves. You don’t need exhaustive preparation anymore. You need methodology and the ability to produce content when it matters. When the right moment arrives, you strike.
Published:
Original Publication Date: March 15, 2026
Published:
Original Publication Date: March 15, 2026
Published:
A PhD system built on a 19th-century apprenticeship model is failing by nearly every empirical measure — ~40–50% attrition, depression rates six times the general population, and tenure-track placement below 15% in many fields — while venture capital has spent four decades perfecting bilateral contracts that manage exactly the risks PhD programs ignore: information asymmetry, moral hazard, hold-up, and misaligned incentives. The literature across economics, education policy, signaling theory, and AI research converges on a striking conclusion: the structural tools to fix the PhD already exist in VC contract design, but academia has never imported them. This research compendium maps the evidentiary landscape across six domains to ground the argument.
Published:
If you browse arXiv regularly, you will be struck by the AI papers whose author lists are so long you have to scroll through multiple pages.
Published:
如果你最近经常逛 arxiv，你一定会被那些作者名单长到需要翻页的 AI 论文给震撼到。
Published:
To be totally honest with you, you don’t need to understand the underlying architecture of large language models to get your name on a top-tier AI paper right now, you just need to be a true expert in your own field.
Published:
坦率的讲，你不需要懂任何大模型底层的原理，只要你是一个真的懂行的领域专家，你现在就能在顶级 AI 论文上署名。
Published:
Doing research is gambling. In fact, for people without money, research is gambling, and so is everything else.
Published:
Research manuscript repository for meteor speed variations study published on Journal of Geophysical Research (JGR): Space Physics.
Published:
这是一个关于流星速度变化研究的论文仓库，相关稿件已在 Journal of Geophysical Research (JGR)发表。
Published:
Dr. Zong-Liang Yang’s Land Environment and Atmospheric Dynamics (LEAD) Group at UT-Austin employs satellite remote sensing, earth system modeling, and high-performance computing to advance understanding of Earth system sciences.
Published:
UT Austin 宗良杨博士（Dr. Zong-Liang Yang）领导的陆地环境与大气动力学研究组（LEAD Group）利用卫星遥感、地球系统建模和高性能计算，推动人类对地球系统科学的理解。
Published:
THIS IS A FAKE BLOG. The content below is fabricated and should not be cited or treated as a real interview or factual record.
Published:
胡晨曦的研究揭示，城市化不只是被动地应对极端天气，它还会主动重塑热带气旋如何向沿海特大城市倾泻暴雨。
Published:
Early-career growth usually comes down to a few repeatable layers: alignment, delivery, visibility, communication, leadership, and trust.
Published:
职业早期的成长，通常可以拆成几个可重复的层次：方向对齐、稳定交付、建立可见度、提升沟通、展现领导力，以及赢得信任。
Published:
Contribute just one research skill and you can become a coauthor of a Nature paper.
Published:
只要你贡献一点科研绝活，就能成为 Nature 文章的合著者。
Published:
If you’re serious about understanding LLMs and AI engineering, Simon Willison’s Newsletter is one of the best resources out there. With over 38,000 subscribers, it provides detailed analysis on AI, LLMs, web engineering, open source, data science, and Python.
Published:
如果你认真想要理解 LLM 和 AI 工程，Simon Willison 的 Newsletter 是最好的资源之一。拥有超过 38,000 名订阅者，它提供关于 AI、LLM、Web 工程、开源、数据科学和 Python 的深度分析。
Published:
A collection of USTC academic resources covering career planning, library access, and research tools.
Published:
一份涵盖生涯规划、图书馆访问与研究工具的中科大学术资源合集。
Published:
Securing up to $150,000 in research funding over three years can define a graduate career, and NASA’s FINESST program is the primary vehicle for that transformation. This guide distills the complex application process into actionable strategies for Earth and Space Science researchers seeking to join the next generation of Future Investigators.
Published:
每年 5 万美元且连续 3 年的科研资助，让 NASA FINESST 项目成为地球与空间科学博士生必须争取的黄金机会。这份指南旨在将复杂的申请流程拆解为可操作的策略，帮助下一代“未来研究员”（Future Investigators）在激烈的竞争中脱颖而出。
Published:
A business jet fell out of the sky over Florida yesterday, and all nine people on board died.
Published:
昨天，一架公务机从奥斯汀起飞，飞到佛罗里达上空掉了下来，机上 9 个人全部遇难。
Published:
A salary can arrive on time every two weeks while your ability to leave the job gets smaller every year.
Published:
工资每两周准时到账，一个人离开这份工作的能力，却可能一年比一年小。
Published:
Doing research is gambling. In fact, for people without money, research is gambling, and so is everything else.
Published:
In the fall of 2015, two things happened simultaneously.
Published:
2015 年秋天，有两件事同时发生了。
Published:
Call it a (10^5) gap if you like, once you put semantic entropy and effective training samples on the board, the difference between internet-scale text and embodied robotics data stops sounding like rhetorical inflation and starts sounding almost conservative. You just have to separate raw physics bits from semantics a learner can actually use.
Published:
雷达现在能订阅了。第十二天，我们加了一个每天更新的 RSS 订阅源。
Published:
Lay out every hyped multiAgent product on a table in front of Steve Jobs and he would probably squint, frown, and ask one question, why on earth are you showing the user any of this.
Published:
把现在市面上所有被吹得天花乱坠的multiAgent产品摆在乔布斯面前，他大概率会皱着眉头反问一句，你为什么要让用户看到这些东西。
alias rm=trash Cannot Stop an AI Agent from rm -rf Published:
I asked Claude Code a narrow question: can I protect this machine from Codex accidentally deleting files forever, just by aliasing rm to trash? The honest answer turned out to be no, for a reason that is not obvious until you actually test it, and the fix ended up being a five-layer setup rather than a one-liner.
alias rm=trash 拦不住 AI agent 的 rm -rf Published:
我问了 Claude Code 一个很具体的问题：能不能只靠把 rm alias 成 trash，来防止 Codex 意外把文件永久删掉。真实的答案是不能，原因不实测根本看不出来，最后落地的也不是一行配置，而是五层防护。
Published:
What Happens When You Just Give People Money? Sam Altman’s UBI Experiment
Published:
At the end of AI’s compute road lies the ultimate solution to energy.
Published:
引言：AI 的算力尽头，是能源的终极解法。
Published:
你好，我是 Koutian。第十四天，我们把雷达的几个老毛病修好了，也让它更像一个会自己干活的工具。
Published:
School rankings are a single number standing in for a multi-dimensional decision. School Evaluator is a small web tool that lets you score a school across the dimensions that actually shape your life there. Try Live Demo
Published:
学校排名只是一个用单一数字去代表多维决策的替代品。School Evaluator 是一款小巧的网页工具，让你能在一个真正决定你在那里生活的维度上为学校打分。试试在线演示
Published:
Scientists sell themselves too.
Published:
科学家也是出来卖的。
Published:
Most of biology is averages. You measure a million cells, you get a mean. Eric Greene’s lab figured out how to actually watch one DNA molecule at a time, repair itself, in real time, and that changes what kind of questions biology can ask.
Published:
大部分生物学是平均值。你测一百万个细胞，得到一个均值。Eric Greene 的实验室却想出了怎么一次实时观看一个 DNA 分子自我修复，而这改变了生物学能提出什么样的问题。
Published:
The promise of AI-automated science is intoxicating: imagine machines that can generate hypotheses, design experiments, and publish papers while we sleep.
Published:
AI 自动化科研的愿景令人陶醉：想象机器能在我们睡觉时生成假设、设计实验、发表论文。然而，尽管关于“AI 科学家”的新闻铺天盖地，我们正在撞上一堵根本性的墙。问题不在于算力或数据集规模，而是某种更深刻的东西，根植于科学发现的本质和信息论之中。
Published:
Most physics undergraduates know Chen Ning Yang as a Nobel laureate and the Y in Yang-Mills. Building him a chronicle made me see something the textbooks gloss over: the most consequential single fact about his life is who his father was, and what that father had set up for him before he was born.
Published:
大多数物理本科生知道杨振宁是诺贝尔奖得主，知道他是 Yang-Mills 里的那个 Yang。而为他建一座年鉴，让我看到了教科书略过的东西：他一生中最有分量的一个事实是谁是他的父亲，以及那个父亲在他出生之前，为他铺好了什么。
Published:
The University of Texas at Austin’s Office of Graduate and Postdoctoral Studies featured my work on LinkedIn on August 26, 2026.
Published:
德克萨斯大学奥斯汀分校研究生与博士后事务办公室于 2026 年 8 月 26 日在 LinkedIn 上专题介绍了我的工作。
Published:
Sometimes a piece of old science fiction stays in your head not because of the gadgets, but because it got the emotional structure of the future right.
Published:
有些旧科幻会一直留在脑子里，不只是因为它写了什么炫酷科技，而是因为它提前抓住了未来的情感结构。
Published:
Python and MATLAB tools for creating publication-quality scientific maps and geographic visualizations.
Published:
用 Python 和 MATLAB 创建适合论文发表的科学地图与地理可视化图件。
Published:
The promise of AI-automated science is intoxicating: imagine machines that can generate hypotheses, design experiments, and publish papers while we sleep.
Published:
AI 自动化科研的愿景令人陶醉：想象机器能在我们睡觉时生成假设、设计实验、发表论文。然而，尽管关于“AI 科学家”的新闻铺天盖地，我们正在撞上一堵根本性的墙。问题不在于算力或数据集规模，而是某种更深刻的东西，根植于科学发现的本质和信息论之中。
Published:
I’ve been playing with mixed-language builds (Fortran calling into C++ and C via iso_c_binding) and wanted a demo that was more interesting than “add two numbers across languages.” So I built fortran-zip-bomb: a small program that generates a genuine ZIP bomb — a small archive that expands into a much larger file on decompression.
Published:
我一直在玩混合语言构建（Fortran 通过 iso_c_binding 调用 C++ 和 C），想要一个比”跨语言把两个数加起来”更有意思的演示。于是我建了 fortran-zip-bomb：一个小程序，生成一个真正的 ZIP 炸弹——一个解压时会扩展成大得多的文件的小压缩包。
Published:
Many Chinese teams building embodied AI or AI applications hit the same real-world problem: the domestic models aren’t enough, and they want to call the APIs of foreign models like OpenAI and Anthropic directly. That’s why a swarm of “relay,” “recharge,” and “one OpenAI-compatible interface for you” businesses has sprung up.
Published:
很多做具身智能、做 AI 应用的中国团队，都会遇到同一个现实问题：国内的模型不够用，想直接调 OpenAI、Anthropic 这些海外模型的 API。于是市面上冒出一堆”中转”“代充”“一个 OpenAI 兼容接口给你用”的生意。
Published:
Should Silicon Valley engineers use OpenClaw?
Published:
硅谷大厂的工程师究竟需不需要用 OpenClaw？
Published:
Exchanging money for time is a privilege I rarely indulge in, but a surprise membership benefit that cleared airport security in ten minutes changed my perspective on friction and market regulation. While a Clear Plus membership normally costs over a hundred dollars annually, obtaining it for free through an Uber membership allowed me to experience a level of efficiency that money can’t always buy—at least not without a well-regulated system behind it.
Published:
用金钱换取时间是一种我很少尝试的“特权”，但这次在奥斯汀机场仅用十分钟便完成安检的经历，让我对效率与市场调节有了新的思考。这份特权并非我主动购买，而是通过 Uber 年费会员赠送的 Clear Plus 获得的——原本需要每年支付一百多美金的服务，在免除门槛后，带给我一种金钱也未必能随时买到的流畅体验。
Published:
I realized I had a serious misunderstanding in my “classes are useless” view.
Published:
第十一天，KW-Bench 的能力分级上线了，社区发布也开始了。
alias rm=trash Cannot Stop an AI Agent from rm -rf Published:
I asked Claude Code a narrow question: can I protect this machine from Codex accidentally deleting files forever, just by aliasing rm to trash? The honest answer turned out to be no, for a reason that is not obvious until you actually test it, and the fix ended up being a five-layer setup rather than a one-liner.
alias rm=trash 拦不住 AI agent 的 rm -rf Published:
我问了 Claude Code 一个很具体的问题：能不能只靠把 rm alias 成 trash，来防止 Codex 意外把文件永久删掉。真实的答案是不能，原因不实测根本看不出来，最后落地的也不是一行配置，而是五层防护。
Published:
In October 2023 I asked ChatGPT an over-engineered question: how do you fry rice so that every grain of rice ends up bonded to egg, with no bare rice grains and no isolated clumps of egg sitting off on their own? I saved that conversation as an HTML file, dropped it in my Downloads folder, and did not look at it again for almost three years. Last week I finally turned it into an actual open-source repo, and going back through the original conversation to build it was more interesting than I expected.
Published:
2023 年 10 月，我问了 ChatGPT 一个过度工程化的问题：怎么炒蛋炒饭，才能让每一粒米都裹上蛋，既没有裸露的白米粒，也没有单独抱团、没沾到米的蛋碎？那次对话我存成了一个 HTML 文件，扔进 Downloads 文件夹，然后差不多三年没再打开过。上周我终于把它整理成了一个真正的开源仓库，回头重读那段 2023 年的对话、把它做成代码的过程，比我预想的有意思得多。
Published:
Should Silicon Valley engineers use OpenClaw?
Published:
硅谷大厂的工程师究竟需不需要用 OpenClaw？
Published:
Large model vendors spend rocket-launch money raising a legion of geeks who only want to buy a two-dollar firecracker just to hear it pop.
Published:
大模型厂商用造火箭的成本，培养出了一群只想买两块钱鞭炮听个响的极客。
Published:
Walking University Avenue in Palo Alto today, I finally saw what many Chinese students get wrong, we treat Silicon Valley as a destination.
Published:
When gold gets too expensive relative to silver, you sell your gold and buy silver. When silver catches up, you swap back. This sounds like folk wisdom, but it’s one of the most time-tested strategies in precious metals.
Published:
Most of biology is averages. You measure a million cells, you get a mean. Eric Greene’s lab figured out how to actually watch one DNA molecule at a time, repair itself, in real time, and that changes what kind of questions biology can ask.
Published:
大部分生物学是平均值。你测一百万个细胞，得到一个均值。Eric Greene 的实验室却想出了怎么一次实时观看一个 DNA 分子自我修复，而这改变了生物学能提出什么样的问题。
Published:
Earlier today I published a piece called “Four Ego Mistakes I Made as a 22-Year-Old Founder.” Then I ran it through Gemini 2.5 Pro as an independent reviewer. Gemini’s verdict was that the essay is itself an ego move. I think Gemini is mostly right.
Published:
今天早些时候我发了一篇叫《22 岁 founder 的四个 ego 错误》。然后我把这篇喂给了 Gemini 2.5 Pro 当独立审稿人。Gemini 的判决是：这篇文章本身就是一个 ego 动作。我觉得 Gemini 大体是对的。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
Decades of Earth-system modeling judgment live in PDFs, mailing lists, and senior researchers’ heads, and none of it is loadable by an AI coding agent. earth-space-ai.org is an attempt to fix that.
Published:
数十年来积累的地球系统建模判断散落在 PDF、邮件列表和资深研究者的头脑里，AI 编程智能体无法直接加载这些知识。earth-space-ai.org 试图解决这个问题。
Published:
At today’s Silicon Valley alumni “Ask a Question” AI meet-up, what I wanted to talk about wasn’t how cool Claude Code is, but how it actually re-slices real work.
Published:
今天在硅谷校友会的「师问」AI 交流活动上，我想讲的不是 Claude Code 多酷，而是它到底怎么把真实工作重新拆开。
Published:
Seizing the gateway to the human mind in the AI era is, at bottom, a secret war over token throughput power.
Published:
抢占 AI 时代的人类心智入口，说到底是一场关于 token 吞吐权力的隐秘战争。
Published:
The SciFi account shared Benchmark Radar on X, carrying the paper beyond the eval community into adjacent audiences.
Published:
SciFi 账号在 X 转发 Benchmark Radar，把论文带出评测圈子，送往相邻受众。
Published:
HuggingPapers mentioned Benchmark Radar on X, carrying arXiv:2609.11115 into the daily paper flow.
Published:
HuggingPapers 在 X 提及 Benchmark Radar，把 arXiv:2609.11115 送进每日论文流。
Published:
A community share carried Benchmark Radar further across X, reader to readers.
Published:
一条社区转发把 Benchmark Radar 带往 X 更远处，读者传读者。
Published:
A Moltbook post carried Benchmark Radar onto the agent internet’s front page — the social feed built for agents, by agents.
Published:
一条 Moltbook 帖子把 Benchmark Radar 送上 agent 互联网的首页——那个为 agent 建、由 agent 用的社交 feed。
Published:
Alexy Khrabrov, who runs Strongly Typed AI News, shared its Benchmark Radar lead story on LinkedIn from his own account, carrying it from the news site into professional networks.
Published:
Strongly Typed AI News 的运营者 Alexy Khrabrov 用个人账号在 LinkedIn 转发了 Benchmark Radar 头条，把它从新闻站送进职业网络。
Published:
A launch-coverage share on LinkedIn carried the Benchmark Radar launch story into professional networks.
Published:
一条 LinkedIn 发布报道转发把 Benchmark Radar 发布故事送进职业网络。
Published:
I ended up in the wrong NEXUS line at Vancouver International Airport. By the rules, the mistake was mine. By design, it was an almost predictable wrong turn.
Published:
Beyond the wayfinding problem that led me into the wrong NEXUS line, my connection at Vancouver International Airport left me with several observations unrelated to signs. I noticed an older-looking mix of people, a departures area that offered almost nowhere to work, and a visible contrast between people in Vancouver and Texas. This post records those observations. The full account of the NEXUS incident is in How I “Reasonably” Ended Up in the Wrong NEXUS Line at YVR.
Published:
A university’s favorite performance is to wait until a person no longer needs a medal, then ceremonially drape them in a gown embroidered with the word “Doctor.”
Published:
大学最擅长的表演，莫过于在一个人不再需要勋章时，隆重地为他披上一件绣着「博士」二字的学位袍。
Published:
What do a 7,169-yuan Chinese animated film and a 483-dollar GameStop short squeeze have in common? More than you would think.
Published:
一部票房 7169 元的国产动画，和一只冲到 483 美元的 GameStop 股票，有什么共同点？比你想象的要多。
Published:
你好，我是 Koutian。第十五天，雷达第一次学会主动对外说话，不再只是等你来翻。
Published:
A few days ago Yann LeCun announced his new role in a LinkedIn post, something almost unthinkable in China.
Published:
While you’re still grinding your heart out on “if you just work hard enough, you can move up,” you may not realize the system was never designed to cultivate you in the first place, it was designed to screen you.
Published:
当你还在为“只要努力就能向上流动”而拼命内卷时，你可能没有意识到，这个系统从一开始就不是为了培养你，而是为了筛选你。
Published:
What Happens When You Just Give People Money? Sam Altman’s UBI Experiment
Published:
直接发钱会怎样？奥特曼的全民基本收入实验
Published:
A laser lightsail accelerating to one percent of light speed turns about one percent of the beam into motion. The other ninety-nine percent leaves as light and does no work.
Published:
激光光帆加速到光速的百分之一，光束里大约只有百分之一变成了运动。剩下的百分之九十九以光的形式离开，不做功。
Published:
The same way text foundation models ate NLP and image models ate computer vision, someone is going to build the foundation model that eats spatial reasoning. Gengchen Mai is one of the people taking that bet seriously, and his SEAI Lab at UT Austin is one of the places it is being built.
Published:
文本基础模型已经改变了 NLP，图像模型也改变了计算机视觉。空间推理领域同样会出现自己的基础模型。Gengchen Mai 正在研究这一方向，他在 UT Austin 的 SEAI Lab 也开展相关工作。
Published:
An elegant couple mood tracking website with self-hosted backend and SQLite, displaying daily mood records in GitHub contribution graph style.
Published:
这是一个优雅的情侣心情记录网站，采用自托管后端和 SQLite，以 GitHub contribution graph 风格展示每日心情记录。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
Hah, I can’t help but laugh. It just hit me: the first time I ever used Markdown was back when I was building a poetry society. And that’s also when I first learned about Git. Looking back now, if you put all the founding members of that poetry society together, you’d almost have two unicorns.
Published:
哎，我他妈笑了。我忽然想起来，我最早用 Markdown，就是之前创建诗社的时候。知道 Git，也是在那个时候。现在回头看，整个诗社的元老凑在一起，真的快有两个独角兽了。
Published:
Many Chinese teams building embodied AI or AI applications hit the same real-world problem: the domestic models aren’t enough, and they want to call the APIs of foreign models like OpenAI and Anthropic directly. That’s why a swarm of “relay,” “recharge,” and “one OpenAI-compatible interface for you” businesses has sprung up.
Published:
很多做具身智能、做 AI 应用的中国团队，都会遇到同一个现实问题：国内的模型不够用，想直接调 OpenAI、Anthropic 这些海外模型的 API。于是市面上冒出一堆”中转”“代充”“一个 OpenAI 兼容接口给你用”的生意。
Published:
Seizing the gateway to the human mind in the AI era is, at bottom, a secret war over token throughput power.
Published:
抢占 AI 时代的人类心智入口，说到底是一场关于 token 吞吐权力的隐秘战争。
Published:
Earlier today I published a piece called “Four Ego Mistakes I Made as a 22-Year-Old Founder.” Then I ran it through Gemini 2.5 Pro as an independent reviewer. Gemini’s verdict was that the essay is itself an ego move. I think Gemini is mostly right.
Published:
今天早些时候我发了一篇叫《22 岁 founder 的四个 ego 错误》。然后我把这篇喂给了 Gemini 2.5 Pro 当独立审稿人。Gemini 的判决是：这篇文章本身就是一个 ego 动作。我觉得 Gemini 大体是对的。
Published:
People love saying that ten years to a billion-dollar valuation is the normal pace for a unicorn, neither fast nor slow. But once you actually crunch the compound numbers, you realize what hides behind that word “normal” is a beast that doubles in value every single year.
Published:
In an era where AI can write code in seconds, I just learned the hard way that blindly moving fast is actually slowing me down.
Published:
在这个闭着眼睛就能让 AI 写代码的时代，我今天结结实实地挨了一锤。
Published:
If you’ve been watching the business map of the AI track lately, you can’t help but be shaken by that trend feverishly racing from “human-machine collaboration” toward “multi-agent autonomy.”
Published:
如果你最近在看AI赛道的商业图谱，你一定会被那个从「人机协作」向「多智能体自治」疯狂演进的趋势给震撼到。
Published:
Lay out every hyped multiAgent product on a table in front of Steve Jobs and he would probably squint, frown, and ask one question, why on earth are you showing the user any of this.
Published:
把现在市面上所有被吹得天花乱坠的multiAgent产品摆在乔布斯面前，他大概率会皱着眉头反问一句，你为什么要让用户看到这些东西。
Published:
This is the worst startup product I have seen.
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
Published:
Yuxuan and I have been discussing AI agent ideas for a while. Yesterday, we finally decided: we’re building Atomize, a task-breaking agent system evolved from Goblin Tools.
Published:
我和 Yuxuan 讨论 AI Agent 的想法已经有一段时间了。昨天，我们终于做出了决定：我们要做 Atomize，一个从 Goblin Tools 演化而来的任务分解 Agent 系统。
Published:
Why do some people who seem to work less than you succeed earlier? Because success is a question of timing first, and ability second.
Published:
You think you picked your major rationally, but the choice itself is a bet.
Published:
On February 1 this year, I wrote down a thought about one question: what kind of company is the greatest company in the world?
Published:
今年 2 月 1 日，我写过一段话，想的是一个问题，什么样的公司是世界上最伟大的公司。
Published:
Secretly recording a customer interview can turn a useful research habit into a privacy problem in one click.
Published:
偷偷录一次客户访谈，可能把一个正常的产品研究流程变成隐私和合规事故。
Published:
Five YC rebels: some were expelled, some were welcomed back. What happened next?
Published:
A bit of history from a casual chat: in the US, science students used to get bigger stipends than business, law or medicine students, because science students were poor.
Published:
Lay out every hyped multiAgent product on a table in front of Steve Jobs and he would probably squint, frown, and ask one question, why on earth are you showing the user any of this.
Published:
把现在市面上所有被吹得天花乱坠的multiAgent产品摆在乔布斯面前，他大概率会皱着眉头反问一句，你为什么要让用户看到这些东西。
Published:
A year and a half into my PhD in the US, the best training I have received is learning to tell a good story.
Published:
The opportunity is not the news itself. The opportunity is the friction created by the news.
Published:
新闻本身不是机会。新闻制造的摩擦才是机会。
Published:
Research, stock trading, and startups all share a common pattern: success comes from combining a defensible core competency with the ability to ride trending waves. You don’t need exhaustive preparation anymore. You need methodology and the ability to produce content when it matters. When the right moment arrives, you strike.
Published:
The signal truly worth heeding is not that AI companies are growing too slowly, but that they are growing fast while cash-flow pressure keeps mounting.
Published:
真正值得警惕的信号不是 AI 公司增长太慢，而是它们增长得很快，现金流压力却仍在放大。
Published:
Why not just go for 20 million, or 30 million?
Published:
How can a generative-AI student society do more than host talks, repost news, and chase the latest model?
Published:
一个生成式 AI 学生社团，怎样才能不止于办讲座、转发新闻和追逐下一款模型？
Published:
UT01 is a free, student-built collection of resources for The University of Texas at Austin. It brings practical campus information and community projects into one public site.
Published:
Sunk cost should never be a reason for a decision, and that includes your choice of major.
Published:
One cup of milk tea, a tip of zero selected on the screen, and the shop still added a mandatory 18%.
Published:
You think you picked your major rationally, but the choice itself is a bet.
Published:
The sense of agency that took me more than twenty years to grow, these seventeen and eighteen year old students got far more easily, just by being at Stanford, being in America, or maybe going to high school in Shanghai first.
Published:
Walking University Avenue in Palo Alto today, I finally saw what many Chinese students get wrong, we treat Silicon Valley as a destination.
Published:
On a stormy Sunday afternoon in 1988, Edmund Clarke, a Carnegie Mellon professor and Turing Award winner, came to lecture at Fudan University and found the room nearly empty.
Published:
You land two summer research interviews at the same time. Is your first reaction joy, or fear of offending both professors?
Published:
A student came to me today for advice on summer research applications, and when I asked what he wanted out of it, he said it straight: a return offer.
Published:
Why not just go for 20 million, or 30 million?
Published:
If all your practice comes from school exams, you will never get that many chances to practice.
Published:
For kids who grew up in China, the first problem in applying abroad, and often the biggest, is not what to do. It is how they think.
Published:
雷达现在能订阅了。第十二天，我们加了一个每天更新的 RSS 订阅源。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
You land two summer research interviews at the same time. Is your first reaction joy, or fear of offending both professors?
Published:
A student came to me today for advice on summer research applications, and when I asked what he wanted out of it, he said it straight: a return offer.
Published:
Sunk cost should never be a reason for a decision, and that includes your choice of major.
Published:
A comprehensive tool built with React, Vite, and Supabase to help individuals plan and navigate their careers in the AI industry.
Published:
这是一个基于 React、Vite 和 Supabase 构建的综合工具，帮助个人规划并导航自己在 AI 行业中的职业发展路径。
Published:
A PhD is hard. A PhD in 2026 — with the field moving faster than your committee can read — is a different kind of hard. AI PhD Survival Guide is a handbook for surviving and finishing an AI or ML PhD without burning out and without falling behind.
Published:
读博很难。2026 年读博——当领域进展快到你的委员会都读不过来时——是另一种难。AI PhD Survival Guide 是一本手册，教你如何在 AI 或机器学习博士研究中活下来并毕业，既不会燃烧殆尽，也不至于掉队。
Published:
Every knowledge worker is now competing with an LLM for some part of their job. AI Survival Guide is a handbook for figuring out which parts, what to do about it, and how to come out the other side better at your craft instead of replaced by it.
Published:
如今，每一位知识工作者都在用某项工作的某些部分与 LLM 竞争。AI Survival Guide 是一本手册，帮你弄清是哪些部分、该怎么应对，以及如何带着更强的技艺从这场竞争的另一头走出来，而不是被它取代。
Published:
The US is full of small traps that nobody warns you about until after you’ve fallen into one. US Survival Guide is a handbook for Chinese students and newcomers on identifying risk early, knowing your rights, and building a financial defense before you need it. Read the Guide
Published:
美国布满了各式各样的暗坑，在你掉进去之前，没人会提醒你。US Survival Guide 是一本写给中国学生和新移民的手册，教你及早识别风险、了解自己的权利，并在你需要之前先筑好财务防线。阅读指南
Published:
I’ve been playing with mixed-language builds (Fortran calling into C++ and C via iso_c_binding) and wanted a demo that was more interesting than “add two numbers across languages.” So I built fortran-zip-bomb: a small program that generates a genuine ZIP bomb — a small archive that expands into a much larger file on decompression.
Published:
我一直在玩混合语言构建（Fortran 通过 iso_c_binding 调用 C++ 和 C），想要一个比”跨语言把两个数加起来”更有意思的演示。于是我建了 fortran-zip-bomb：一个小程序，生成一个真正的 ZIP 炸弹——一个解压时会扩展成大得多的文件的小压缩包。
Published:
The roar of cooling systems in the Texas Advanced Computing Center (TACC) isn’t just noise—it’s the power of supercomputers translating the overwhelming dimensionality of climate data into something students can finally see and understand.
Published:
德州高级计算中心（TACC）里冷却系统的轰鸣不只是噪音——那是超级计算机的力量，正在把气候数据那令人不知所措的多维性，转化成学生最终能看见、能理解的东西。
Published:
The promise of AI-automated science is intoxicating: imagine machines that can generate hypotheses, design experiments, and publish papers while we sleep.
Published:
AI 自动化科研的愿景令人陶醉：想象机器能在我们睡觉时生成假设、设计实验、发表论文。然而，尽管关于“AI 科学家”的新闻铺天盖地，我们正在撞上一堵根本性的墙。问题不在于算力或数据集规模，而是某种更深刻的东西，根植于科学发现的本质和信息论之中。
Published:
I have a theory that the way our bodies are built is adapted to the ancient world.
Published:
When something urgent lands at work and there is no time to prepare, so you stay up all night, that is a black swan.
Published:
The sharpest minds keep moving into AI labs, and it isn’t just for the money.
Published:
这是《颠覆性创新者的思维模式》的五个月后续研究。
Published:
In April 2024 I was the only undergraduate invited to speak among more than 40 scholars.
Published:
On a stormy Sunday afternoon in 1988, Edmund Clarke, a Carnegie Mellon professor and Turing Award winner, came to lecture at Fudan University and found the room nearly empty.
Published:
I spent a whole evening with Gemini taking apart the Buy Borrow Die playbook that American billionaires run, expecting that with a little scaling down I could just copy the moves, and what I found instead was that almost every single move has an F-1 trapdoor underneath it, and the list of things I can actually do fits on one page.
Published:
Research date: 2026-05-26
Published:
研究时间：2026-05-26
Published:
@elonmusk, asking for a friend, has Tesla FSD been tested in weather like this?
Published:
The silence surrounding the Alamo chapel in San Antonio belies the brutal, thirteen-day siege that transformed this former Spanish mission into the ultimate symbol of Texan independence. Standing before its weathered facade today, one can almost hear the echoes of a conflict that remains one of the most legendary chapters in American history.
Published:
Port Aransas South Jetty published an Institute Insights feature on May 28, 2026 about how our University of Texas team is using satellite remote sensing and AI to monitor Texas coastal water quality.
Published:
阿兰瑟斯港南防波堤 于 2026 年 5 月 28 日发表了一篇 Institute Insights 专题，讲的是我们德克萨斯大学团队如何用卫星遥感与 AI 监测德克萨斯沿海水质。
Published:
Hah, I can’t help but laugh. It just hit me: the first time I ever used Markdown was back when I was building a poetry society. And that’s also when I first learned about Git. Looking back now, if you put all the founding members of that poetry society together, you’d almost have two unicorns.
Published:
哎，我他妈笑了。我忽然想起来，我最早用 Markdown，就是之前创建诗社的时候。知道 Git，也是在那个时候。现在回头看，整个诗社的元老凑在一起，真的快有两个独角兽了。
Published:
People love saying that ten years to a billion-dollar valuation is the normal pace for a unicorn, neither fast nor slow. But once you actually crunch the compound numbers, you realize what hides behind that word “normal” is a beast that doubles in value every single year.
Published:
Top physics researchers aren’t blind to what AI is doing in the application layer. They’re locked by their own evaluation system and can’t afford to look.
Published:
In an era where AI can write code in seconds, I just learned the hard way that blindly moving fast is actually slowing me down.
Published:
在这个闭着眼睛就能让 AI 写代码的时代，我今天结结实实地挨了一锤。
Published:
Sometimes a piece of old science fiction stays in your head not because of the gadgets, but because it got the emotional structure of the future right.
Published:
有些旧科幻会一直留在脑子里，不只是因为它写了什么炫酷科技，而是因为它提前抓住了未来的情感结构。
Published:
Every generation has its own way of managing work: paper notebooks, SaaS task managers, and now programmable agentic workflows powered by tools like OpenClaw heartbeat.
Published:
每一代人都有自己管理任务的方式：最早是纸和笔，后来是 SaaS 任务管理工具，现在则开始进入像 OpenClaw heartbeat 这样可程序化、可持续运行的 agentic workflow 时代。
Published:
I spent a whole evening with Gemini taking apart the Buy Borrow Die playbook that American billionaires run, expecting that with a little scaling down I could just copy the moves, and what I found instead was that almost every single move has an F-1 trapdoor underneath it, and the list of things I can actually do fits on one page.
Published:
When gold gets too expensive relative to silver, you sell your gold and buy silver. When silver catches up, you swap back. This sounds like folk wisdom, but it’s one of the most time-tested strategies in precious metals.
| The Mindset of Disruptive Innovators | What Did the People Who Changed the World Actually See? | An Epistemic Crisis in an Age of Abundance |
Published:
What do the people who change the world see that ordinary people can’t?
Published:
改变世界的人到底看到了什么是普通人看不到的？
Published:
A hand-controlled 3D visualization of your AI conversation history - fly through ChatGPT conversations and explore artificial intelligence concepts. Try Live Demo
Published:
一个支持手势控制的 3D 可视化项目，把你的 AI 对话历史变成可以飞行探索的星系。你可以在 ChatGPT 对话之间穿梭，也可以把抽象的人工智能概念变成一个能亲手操作的可视化空间。Try Live Demo
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
Why AI agents lack temporal awareness, and how injecting timestamps transforms them from stateless automata into coherent entities.
Published:
TQQQ is the 3x-leveraged Nasdaq-100 ETF. It is also one of the most asymmetric instruments retail investors touch — the upside is real, the drawdowns are brutal, and the daily-rebalance math means buy-and-hold doesn’t behave the way most people assume. TQQQ ML Trend is an experiment in using machine learning to predict the trend regime, not the price.
Published:
TQQQ 是纳斯达克 100 指数的 3 倍杠杆 ETF。它也是散户接触过的最不对称的金融工具之一——上行空间真实存在，回撤却十分残酷，而每日再平衡的数学逻辑意味着”买入并持有”并不像大多数人以为的那样运转。TQQQ ML Trend 是一个用机器学习来预测趋势状态的实验，而不是去预测价格。
Published:
A student came to me today for advice on summer research applications, and when I asked what he wanted out of it, he said it straight: a return offer.
Published:
One cup of milk tea, a tip of zero selected on the screen, and the shop still added a mandatory 18%.
Published:
The Tmux Orchestrator enables Claude agents to work autonomously, schedule their own check-ins, and coordinate across multiple projects without human intervention - a project I explored and learned a lot from.
Published:
Tmux Orchestrator 让 Claude agents 能够自主工作、自己安排 check-in，并在多个项目之间协同推进——这是一个我探索过、也从中学到很多的项目。
Published:
Every indie developer who has ever burned through a free API tier knows the feeling: you build something cool, and then your OPENAI_API_KEY runs out at the worst possible moment. hao-tokens is a curated list of the legitimate ways to get free or low-cost LLM tokens so you can keep building.
Published:
每一个曾经把某个免费 API 层额度烧光的独立开发者都知道那种感觉：你做了很酷的东西，然后你的 OPENAI_API_KEY 在最不该耗尽的时候用光了。hao-tokens 是一份经过整理、合法获取免费或低价 LLM token 的清单，让你能继续开发下去。
Published:
TQQQ is the 3x-leveraged Nasdaq-100 ETF. It is also one of the most asymmetric instruments retail investors touch — the upside is real, the drawdowns are brutal, and the daily-rebalance math means buy-and-hold doesn’t behave the way most people assume. TQQQ ML Trend is an experiment in using machine learning to predict the trend regime, not the price.
Published:
TQQQ 是纳斯达克 100 指数的 3 倍杠杆 ETF。它也是散户接触过的最不对称的金融工具之一——上行空间真实存在，回撤却十分残酷，而每日再平衡的数学逻辑意味着”买入并持有”并不像大多数人以为的那样运转。TQQQ ML Trend 是一个用机器学习来预测趋势状态的实验，而不是去预测价格。
Published:
TQQQ is the 3x-leveraged Nasdaq-100 ETF. It is also one of the most asymmetric instruments retail investors touch — the upside is real, the drawdowns are brutal, and the daily-rebalance math means buy-and-hold doesn’t behave the way most people assume. TQQQ ML Trend is an experiment in using machine learning to predict the trend regime, not the price.
Published:
TQQQ 是纳斯达克 100 指数的 3 倍杠杆 ETF。它也是散户接触过的最不对称的金融工具之一——上行空间真实存在，回撤却十分残酷，而每日再平衡的数学逻辑意味着”买入并持有”并不像大多数人以为的那样运转。TQQQ ML Trend 是一个用机器学习来预测趋势状态的实验，而不是去预测价格。
Published:
When gold gets too expensive relative to silver, you sell your gold and buy silver. When silver catches up, you swap back. This sounds like folk wisdom, but it’s one of the most time-tested strategies in precious metals.
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
Why AI agents lack temporal awareness, and how injecting timestamps transforms them from stateless automata into coherent entities.
Published:
为什么AI Agent缺乏时间意识，以及注入时间戳如何将它们从无状态自动机转变为连贯的实体。
Published:
A receipt in Singapore showed S$5.50, while a U.S. credit-card account displayed only US$4.27. Another purchase of about S$48 appeared as roughly US$38. A bank-card transit ride also failed to appear immediately as pending.
Published:
在新加坡消费时，一张收据写着 S$5.50，但美国信用卡账户只显示 US$4.27。另一笔约 S$48 的消费，则显示为约 US$38。与此同时，刷银行卡乘坐公共交通后，交易也没有立刻出现在 pending 中。
Published:
The thing that almost stranded me on this international trip was not a typhoon, but several unrelated systems deciding at the same time that they did not trust me.
Published:
这次国际旅行真正差点把我留在原地的，不是台风，而是几个互不认识的系统同时决定不相信我。
Published:
The silence surrounding the Alamo chapel in San Antonio belies the brutal, thirteen-day siege that transformed this former Spanish mission into the ultimate symbol of Texan independence. Standing before its weathered facade today, one can almost hear the echoes of a conflict that remains one of the most legendary chapters in American history.
Published:
At the Metropolitan Museum of Art in New York, I walked into a real ancient Egyptian tomb, not a replica or a model, but a tomb that stood in the Egyptian desert more than four thousand years ago and was moved whole to Manhattan.
Published:
Exchanging money for time is a privilege I rarely indulge in, but a surprise membership benefit that cleared airport security in ten minutes changed my perspective on friction and market regulation. While a Clear Plus membership normally costs over a hundred dollars annually, obtaining it for free through an Uber membership allowed me to experience a level of efficiency that money can’t always buy—at least not without a well-regulated system behind it.
Published:
用金钱换取时间是一种我很少尝试的“特权”，但这次在奥斯汀机场仅用十分钟便完成安检的经历，让我对效率与市场调节有了新的思考。这份特权并非我主动购买，而是通过 Uber 年费会员赠送的 Clear Plus 获得的——原本需要每年支付一百多美金的服务，在免除门槛后，带给我一种金钱也未必能随时买到的流畅体验。
Published:
On May 17, 2025, I had about nine hours alone in Hong Kong between flights and trains, and I planned them around Victoria Harbour and the city’s richest commercial districts.
Published:
2025年5月17日，我一人途经香港，可游览时间约九个小时，目标是维多利亚港和香港最繁华的商业地段。
Published:
The clearest advice I recently received from a friend who has visited Japan several times was simple: plan the trip first.
Published:
The opportunity is not the news itself. The opportunity is the friction created by the news.
Published:
THIS IS A FAKE BLOG. The content below is fabricated and should not be cited or treated as a real interview or factual record.
Published:
胡晨曦的研究揭示，城市化不只是被动地应对极端天气，它还会主动重塑热带气旋如何向沿海特大城市倾泻暴雨。
Published:
I plugged a Western Digital external drive full of data into a MacBook Air, tried to copy a file onto it, and nothing happened. No error dialog, no progress bar, just a drive that would let me read everything and write nothing. My first instinct was that something was broken, or that I needed to fix permissions. Both were wrong, and chasing the wrong explanation almost led me to permanently downgrade the security of the whole laptop.
Published:
我把一块装满数据的西部数据（WD）移动硬盘插到 MacBook Air 上，想往里拷一个文件，结果什么都没发生。没有报错弹窗，没有进度条，就是一块能读出所有东西、却一个字节都写不进去的硬盘。我的第一反应是它坏了，或者是我得去修一下权限。这两个判断都是错的，而且顺着错误的解释找下去，差点让我把整台笔记本的安全性永久降级。
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
You land two summer research interviews at the same time. Is your first reaction joy, or fear of offending both professors?
Published:
Most crypto dashboards either drown you in numbers or hide the signal behind a paywall. Crypto Dashboard is my attempt at a clean, AI-augmented view of the market that runs entirely in your browser. Try Live Demo
Published:
大多数加密货币仪表盘要么把你淹没在数字里，要么把信号藏在付费墙后面。Crypto Dashboard 是我对一种简洁、由 AI 增强的市场视图的尝试，而且它完全运行在你的浏览器里。试试在线演示
Published:
School rankings are a single number standing in for a multi-dimensional decision. School Evaluator is a small web tool that lets you score a school across the dimensions that actually shape your life there. Try Live Demo
Published:
学校排名只是一个用单一数字去代表多维决策的替代品。School Evaluator 是一款小巧的网页工具，让你能在一个真正决定你在那里生活的维度上为学校打分。试试在线演示
Published:
Exchanging money for time is a privilege I rarely indulge in, but a surprise membership benefit that cleared airport security in ten minutes changed my perspective on friction and market regulation. While a Clear Plus membership normally costs over a hundred dollars annually, obtaining it for free through an Uber membership allowed me to experience a level of efficiency that money can’t always buy—at least not without a well-regulated system behind it.
Published:
用金钱换取时间是一种我很少尝试的“特权”，但这次在奥斯汀机场仅用十分钟便完成安检的经历，让我对效率与市场调节有了新的思考。这份特权并非我主动购买，而是通过 Uber 年费会员赠送的 Clear Plus 获得的——原本需要每年支付一百多美金的服务，在免除门槛后，带给我一种金钱也未必能随时买到的流畅体验。
Published:
What Happens When You Just Give People Money? Sam Altman’s UBI Experiment
Published:
直接发钱会怎样？奥特曼的全民基本收入实验
Published:
A simple, clean HTML page displaying email verification success message with responsive design and professional styling.
Published:
这是一个简洁干净的 HTML 页面，用于展示邮箱验证成功提示，具备响应式布局和专业风格设计。
Published:
For kids who grew up in China, the first problem in applying abroad, and often the biggest, is not what to do. It is how they think.
Published:
People love saying that ten years to a billion-dollar valuation is the normal pace for a unicorn, neither fast nor slow. But once you actually crunch the compound numbers, you realize what hides behind that word “normal” is a beast that doubles in value every single year.
Published:
The US is full of small traps that nobody warns you about until after you’ve fallen into one. US Survival Guide is a handbook for Chinese students and newcomers on identifying risk early, knowing your rights, and building a financial defense before you need it. Read the Guide
Published:
美国布满了各式各样的暗坑，在你掉进去之前，没人会提醒你。US Survival Guide 是一本写给中国学生和新移民的手册，教你及早识别风险、了解自己的权利，并在你需要之前先筑好财务防线。阅读指南
Published:
What Happens When You Just Give People Money? Sam Altman’s UBI Experiment
Published:
THIS IS A FAKE BLOG. The content below is fabricated and should not be cited or treated as a real interview or factual record.
Published:
胡晨曦的研究揭示，城市化不只是被动地应对极端天气，它还会主动重塑热带气旋如何向沿海特大城市倾泻暴雨。
Published:
A collection of USTC academic resources covering career planning, library access, and research tools.
Published:
一份涵盖生涯规划、图书馆访问与研究工具的中科大学术资源合集。
Published:
The University of Texas at Austin’s Office of Graduate and Postdoctoral Studies featured my work on LinkedIn on August 26, 2026.
Published:
德克萨斯大学奥斯汀分校研究生与博士后事务办公室于 2026 年 8 月 26 日在 LinkedIn 上专题介绍了我的工作。
Published:
How can a generative-AI student society do more than host talks, repost news, and chase the latest model?
Published:
一个生成式 AI 学生社团，怎样才能不止于办讲座、转发新闻和追逐下一款模型？
Published:
Port Aransas South Jetty published an Institute Insights feature on May 28, 2026 about how our University of Texas team is using satellite remote sensing and AI to monitor Texas coastal water quality.
Published:
阿兰瑟斯港南防波堤 于 2026 年 5 月 28 日发表了一篇 Institute Insights 专题，讲的是我们德克萨斯大学团队如何用卫星遥感与 AI 监测德克萨斯沿海水质。
Published:
The talk that stayed with me at the Texas Symposium on Machine Learning, Responsible AI & Robotics 2026 was about extending language models, not just making them bigger.
Published:
AlphaEarthHack is the project our team built for the UT Austin Geoscience Hackathon. The goal: see how far we could push AI on Earth system data inside a single weekend. Try Live Demo
Published:
AlphaEarthHack 是我们团队为德州大学奥斯汀分校地球科学黑客松打造的项目。目标只有一个：看看在一个周末内，我们能在地球系统数据上把 AI 推到多远。试试在线演示
Published:
UT01 is a free, student-built collection of resources for The University of Texas at Austin. It brings practical campus information and community projects into one public site.
Published:
Most people think of summer vacation as time off. I started wondering whether that framing undersells it. Summer Calculator is a small web tool that estimates the full value of your summer — money, learning, relationships, health — not just the days you didn’t work. Try Live Demo
Published:
大多数人把暑假当作休息时间。我开始怀疑这种框架是否低估了它。Summer Calculator 是一款小巧的网页工具，用来估算你暑假的完整价值——金钱、学习、人际关系、健康——而不仅仅是你没工作的那些日子。试试在线演示
Published:
People love saying that ten years to a billion-dollar valuation is the normal pace for a unicorn, neither fast nor slow. But once you actually crunch the compound numbers, you realize what hides behind that word “normal” is a beast that doubles in value every single year.
Published:
Here is a view I hold now: whatever you do, you look a lot like a speculator in a market.
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
Published:
Most land surface models treat a tree like a passive straw. Water comes in at the roots, water leaves at the leaves, end of story. Ashley Matheny’s research basically says no, a tree is an active hydraulic system with storage, capacitance, and a strategy, and if you do not model it that way you are going to be wrong about drought.
Published:
大多数陆地表面模型都把一棵树当成一根被动的吸管。水从根进来，水从叶出去，故事就这么简单。Ashley Matheny 的研究基本上在说：不对，树是一个带有储水、电容和策略的活跃水力系统，如果你不这样建模，你在干旱问题上就会犯错误。
Published:
Doing research is gambling. In fact, for people without money, research is gambling, and so is everything else.
Published:
Virtual cell models powered by neuromorphic computing and gene language models represent one of the most capital-intensive and scientifically ambitious convergences in biotech AI. If you’re evaluating this space as a VC intern, you need to understand three core components: what these systems actually do, why the market is moving now, and where the investable opportunities lie.
Published:
Original Publication Date: March 15, 2026
Published:
Original Publication Date: March 15, 2026
Published:
A PhD system built on a 19th-century apprenticeship model is failing by nearly every empirical measure — ~40–50% attrition, depression rates six times the general population, and tenure-track placement below 15% in many fields — while venture capital has spent four decades perfecting bilateral contracts that manage exactly the risks PhD programs ignore: information asymmetry, moral hazard, hold-up, and misaligned incentives. The literature across economics, education policy, signaling theory, and AI research converges on a striking conclusion: the structural tools to fix the PhD already exist in VC contract design, but academia has never imported them. This research compendium maps the evidentiary landscape across six domains to ground the argument.
Published:
Large model vendors spend rocket-launch money raising a legion of geeks who only want to buy a two-dollar firecracker just to hear it pop.
Published:
大模型厂商用造火箭的成本，培养出了一群只想买两块钱鞭炮听个响的极客。
Published:
At today’s Silicon Valley alumni “Ask a Question” AI meet-up, what I wanted to talk about wasn’t how cool Claude Code is, but how it actually re-slices real work.
Published:
今天在硅谷校友会的「师问」AI 交流活动上，我想讲的不是 Claude Code 多酷，而是它到底怎么把真实工作重新拆开。
Published:
In an era where AI can write code in seconds, I just learned the hard way that blindly moving fast is actually slowing me down.
Published:
在这个闭着眼睛就能让 AI 写代码的时代，我今天结结实实地挨了一锤。
Published:
Large model vendors spend rocket-launch money raising a legion of geeks who only want to buy a two-dollar firecracker just to hear it pop.
Published:
大模型厂商用造火箭的成本，培养出了一群只想买两块钱鞭炮听个响的极客。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
I contributed to the startup documentary Starting Point One. The film records early-stage builders and the choices behind starting something from zero.
Published:
我参与了创业纪录片《Starting Point One》。这部片子记录早期创业者，也记录他们从零开始时做出的选择。
Published:
I appeared in Episode 1 of Mango TV’s Gold Rush 2026 (《夺金2026》) as a young founder working in Austin. The segment offers a brief television snapshot of the projects and questions I was pursuing at the time.
Published:
我在芒果 TV《夺金2026》第一集中以奥斯汀青年创业者的身份出镜。这个片段很短，记录了当时我正在推进的项目和思考的问题。
Published:
Virtual cell models powered by neuromorphic computing and gene language models represent one of the most capital-intensive and scientifically ambitious convergences in biotech AI. If you’re evaluating this space as a VC intern, you need to understand three core components: what these systems actually do, why the market is moving now, and where the investable opportunities lie.
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
The roar of cooling systems in the Texas Advanced Computing Center (TACC) isn’t just noise—it’s the power of supercomputers translating the overwhelming dimensionality of climate data into something students can finally see and understand.
Published:
德州高级计算中心（TACC）里冷却系统的轰鸣不只是噪音——那是超级计算机的力量，正在把气候数据那令人不知所措的多维性，转化成学生最终能看见、能理解的东西。
Published:
A hand-controlled 3D visualization of your AI conversation history - fly through ChatGPT conversations and explore artificial intelligence concepts. Try Live Demo
Published:
一个支持手势控制的 3D 可视化项目，把你的 AI 对话历史变成可以飞行探索的星系。你可以在 ChatGPT 对话之间穿梭，也可以把抽象的人工智能概念变成一个能亲手操作的可视化空间。Try Live Demo
Published:
Python and MATLAB tools for creating publication-quality scientific maps and geographic visualizations.
Published:
用 Python 和 MATLAB 创建适合论文发表的科学地图与地理可视化图件。
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
Wispr Flow is the best AI product I’ve used this year. But I haven’t paid them a cent. Here’s why that’s both impressive and concerning.
Published:
Wispr Flow 是我今年用过的最好的AI产品。但我一分钱都没付给他们。这既令人印象深刻，又令人担忧。
| Warren Buffett: The “Open Source” Leader Half a Century Before GitHub | 巴菲特：比GitHub早了半个世纪的“开源”运动领袖 |
Published:
In an era driven by code, collaboration, and transparency, GitHub has become synonymous with the “open source” spirit. But if we turn our gaze to the world of finance, we find a pioneer of “open source” who predates the birth of GitHub by a full half-century. His name is Warren Buffett.
Published:
在当今这个由代码、协作和透明度驱动的时代，GitHub 成为了”开源”精神的代名词。但如果我们将目光投向金融界，会发现一位”开源”的先行者，他比 GitHub 的诞生早了整整半个世纪。他就是沃伦·巴菲特。
Published:
I ended up in the wrong NEXUS line at Vancouver International Airport. By the rules, the mistake was mine. By design, it was an almost predictable wrong turn.
Published:
Research date: 2026-05-26
Published:
研究时间：2026-05-26
Published:
@elonmusk, asking for a friend, has Tesla FSD been tested in weather like this?
Published:
An elegant couple mood tracking website with self-hosted backend and SQLite, displaying daily mood records in GitHub contribution graph style.
Published:
这是一个优雅的情侣心情记录网站，采用自托管后端和 SQLite，以 GitHub contribution graph 风格展示每日心情记录。
Published:
“这b人生过的值不值” — roughly, “was this damn life worth it?” — is a phrase that does a lot of emotional work on the Chinese internet. ZuoGuoHuaDiao (做过划掉) is a small web tool that takes the phrase seriously and tries to answer it with numbers. Try Live Demo
Published:
“这b人生过的值不值”——简单说，就是”这该死的一生过得值不值？”——这句在中国互联网上承载了大量情感的话语。ZuoGuoHuaDiao（做过划掉）是一款小巧的网页工具，它认真对待这句话，并试图用数字来回答它。试试在线演示
Published:
Most people think of summer vacation as time off. I started wondering whether that framing undersells it. Summer Calculator is a small web tool that estimates the full value of your summer — money, learning, relationships, health — not just the days you didn’t work. Try Live Demo
Published:
大多数人把暑假当作休息时间。我开始怀疑这种框架是否低估了它。Summer Calculator 是一款小巧的网页工具，用来估算你暑假的完整价值——金钱、学习、人际关系、健康——而不仅仅是你没工作的那些日子。试试在线演示
Published:
A web-based platform for facilitating cross-cultural communication and understanding.
Published:
一个基于 Web 的平台，用于促进跨文化交流与理解。
Published:
UT01 is a free, student-built collection of resources for The University of Texas at Austin. It brings practical campus information and community projects into one public site.
Published:
A simple, clean HTML page displaying email verification success message with responsive design and professional styling.
Published:
这是一个简洁干净的 HTML 页面，用于展示邮箱验证成功提示，具备响应式布局和专业风格设计。
Published:
A handcrafted website universe where every webpage, balloon, click, and line of code says “I love you” - designed exclusively for Momo. Visit Live Site
Published:
这是一个手工打造的网站小宇宙：每一个网页、气球、点击和每一行代码都在说“我爱你”——只为 Momo 而设计。访问在线站点
Published:
Today I had an interesting experience collaborating with Claude Code to completely overhaul my personal academic website. As a PhD student in Geological and Earth Sciences at UT Austin, I needed to update my GitHub Pages site with real professional information instead of the placeholder content that had been sitting there.
Published:
今天我经历了一次挺有意思的合作：和 Claude Code 一起，把我的个人学术网站彻底重做了一遍。作为 UT Austin Geological and Earth Sciences 的博士生，我需要把 GitHub Pages 站点从一堆占位符内容，更新成真正能代表我专业背景的信息。
Published:
From the vast datasets of Earth System Models to the specialized niches of high-performance computing, this space documents my journey as a PhD student at UT Austin pushing the boundaries of Geological and Earth Sciences. This website is more than just a portfolio—it’s a hub where data-driven climate science meets the practical challenges of modern research.
Published:
在 UT Austin 攻读地球科学博士学位的过程中，我始终在试图寻找复杂气候模型与真实世界影响之间的联结。这个学术网站不仅是我研究、论文和项目经历的展示入口，更是我记录如何利用数据驱动的科学方法去理解地球系统演变的思考空间。
Published:
你好，我是 Koutian。第十五天，雷达第一次学会主动对外说话，不再只是等你来翻。
Published:
I plugged a Western Digital external drive full of data into a MacBook Air, tried to copy a file onto it, and nothing happened. No error dialog, no progress bar, just a drive that would let me read everything and write nothing. My first instinct was that something was broken, or that I needed to fix permissions. Both were wrong, and chasing the wrong explanation almost led me to permanently downgrade the security of the whole laptop.
Published:
我把一块装满数据的西部数据（WD）移动硬盘插到 MacBook Air 上，想往里拷一个文件，结果什么都没发生。没有报错弹窗，没有进度条，就是一块能读出所有东西、却一个字节都写不进去的硬盘。我的第一反应是它坏了，或者是我得去修一下权限。这两个判断都是错的，而且顺着错误的解释找下去，差点让我把整台笔记本的安全性永久降级。
Published:
The most important measure of work may be the energy it consumes, more than the time it takes.
Published:
衡量工作效率，最重要的指标或许不是投入的时间，而是消耗的能源。
Published:
What I actually want to do today is pull myself out from under a mountain of repetitive work.
Published:
我今天真正想做的事情，其实是把自己从一座 mountains of 重复性的 work 里面捞出来。
Published:
To use OpenClaw or not to use OpenClaw, that is the question.
Published:
用不用 OpenClaw？这是个问题
Published:
Every generation has its own way of managing work: paper notebooks, SaaS task managers, and now programmable agentic workflows powered by tools like OpenClaw heartbeat.
Published:
每一代人都有自己管理任务的方式：最早是纸和笔，后来是 SaaS 任务管理工具，现在则开始进入像 OpenClaw heartbeat 这样可程序化、可持续运行的 agentic workflow 时代。
Published:
While doing career planning for my girlfriend, I generated this report. Reading it as I went, I found it rather surreal.
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
While doing career planning for my girlfriend, I generated this report. Reading it as I went, I found it rather surreal.
Published:
帮女朋友做职业规划的时候，我生成了这份报告。一边看一边觉得，挺魔幻的。
Published:
Earth system models are some of the most complex scientific software ever written, and they are also some of the worst-documented for newcomers. I have been building a series of “skill packages” — structured, progressive-disclosure knowledge bundles — for the major Earth system and land surface models, designed to be used by both new graduate students and AI coding agents.
Published:
地球系统模型是人类写过的、有史以来最复杂的一批科学软件，可它们对新手来说偏偏又是文档最糟糕的一批。我一直在为主要的几大地球系统和陆地表面模型构建一系列”技能包”——结构化的、渐进式披露的知识包——设计给刚入门的研究生和 AI 编码代理两类使用者使用。
Published:
What I actually want to do today is pull myself out from under a mountain of repetitive work.
Published:
我今天真正想做的事情，其实是把自己从一座 mountains of 重复性的 work 里面捞出来。
Published:
When an AI writing assistant turns a paper’s incoherent rambling into a “well-structured pile of something,” we have to admit: it is basically a cyber laxative.
Published:
当AI写作助手把学术论文里的“狗屁不通”梳理成“结构完整的一大坨”时，我们不得不承认：这简直是一剂赛博开塞露。
Published:
Ask which campus actually sits on the best multimodal feedstock for GPT-4V-class perception, Sora-class video, Gemini-scale bundles, and Meta Emu-style image stacks, and the short answer is almost vulgar in how cleanly it splits three big piles. Meta still pulls ahead by a chasm on stills, ByteDance owns the high-velocity short-video river that is really motion plus audio, and X ships the smallest absolute media volume yet the weirdest leverage on tight text-image coupling and live-event semantics.
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
In April 2024 I was the only undergraduate invited to speak among more than 40 scholars.
Published:
A few days ago Yann LeCun announced his new role in a LinkedIn post, something almost unthinkable in China.
Published:
Five YC rebels: some were expelled, some were welcomed back. What happened next?
Published:
五位 YC 叛逆者：有人被驱逐、有人被重新接纳，后来又发生了什么？
Published:
一篇新论文把 1901 到 2023 年所有 739 位科学类诺贝尔奖得主从童年扒到现在，得出了一个特别狠的数字，按目前的进步速度，一个出生在低收入国家的孩子要等大约 600 年，才能拿到和富裕国家孩子一样的诺奖机会。
Published:
在中国长大的孩子，留学申请上遇到的第一个问题，往往也是最大的问题，不是做什么事儿，而是在想什么东西。
Published:
我曾经以为，自己因为好玩在 LinkedIn 上手动添加几千个人，是假把式，没有任何用。
Published:
前几天杨立昆（Yann LeCun）在 LinkedIn 上发帖，说自己担任了一个新的职位，这在国内是几乎不可以想象的。
Published:
从来没有什么最顶级的、普遍适用的方法论，只有一个，就是在一个行业里面深耕，建立个人影响力，建立个人品牌。
Published:
我刚刚在想，为什么我对证券的客服还挺客气，对中介却不能客气。
Published:
我用了二十多年才培育起来的主体性，这些十七八岁的留子，就因为在 Stanford，在美国，或许更早在上海读了高中，就轻易许多地拿到了。
Published:
你的下一台云服务器是活的，每三天要喂一次饭，而且不管你续不续费，它六个月后都会死。
Published:
金涨多了就换成银，银涨多了再换回金，这个听起来像民间偏方的操作，其实是贵金属市场里最经典的增厚策略之一。
Published:
On a United Airlines flight from San Francisco SFO to Shanghai Pudong PVG, the plane had started its descent and the cabin announcements kept reminding everyone to fasten their seatbelts. Outside was night; inside, people were already lit up with the excitement of “finally going home.” It was on this flight that I met a friend.
Published:
在联合航空从旧金山 SFO 飞往上海浦东 PVG 的航班上，飞机已经开始下降，机舱广播一遍遍提醒大家系好安全带。窗外是夜色，舱内的人却已经被“终于回国了”的兴奋点亮。我就在这趟航班上认识了一位朋友。
Published:
I’ve been discussing this question with a friend lately. He wants to use AI for matching in the dating market; I think with a small sample size this is perfectly feasible, you don’t even need to build any search/ads/rec system at all. Just ask the large language model directly, toss in a few users’ profiles, and have it rank them; the whole process is very simple.
Published:
我和朋友最近在讨论这个问题。朋友说想拿 AI 来做婚恋市场的匹配，我觉得在样本量小的情况下，这完全可行——甚至根本不需要搭任何搜广推系统。你直接问大语言模型，把几个用户的简历丢进去，让它排序就好了，整个过程非常简单。
Published:
五位 YC 叛逆者：有人被驱逐、有人被重新接纳，后来又发生了什么？
Published:
这是我见过最糟糕的创业产品。
Published:
这是我见过最糟糕的创业产品。
Published:
Anthropic’s most ruthless move isn’t polishing every product to perfection before release, but daring to throw out a research preview first, then rapidly getting stronger in front of everyone.
Published:
Anthropic 最狠的地方，不是每次都把产品打磨到完美才发布，而是敢先把一个 research preview 扔出来，然后当着所有人的面快速变强。
Published:
我走错了温哥华机场的 NEXUS 通道；规则上是我的错，设计上却是一条几乎可以预测的错误路径。
Published:
2015 年秋天，有两件事同时发生了。
Published:
一份由 AI 制定的完美旅行计划，最后可能会死在一个不存在的“重试”按钮上。
Published:
The “Alumni Real-Life Archive” documents how alumni made decisions at turning points, giving current students a wider range of real paths to consider.
Published:
欢迎加入”校友真实档案”。这里记录的不是成功经验，而是人在关键节点如何做选择。我们希望通过这些真实路径的分享，为在校学生呈现人生的多样性。
Published:
我曾经以为，自己因为好玩在 LinkedIn 上手动添加几千个人，是假把式，没有任何用。
Published:
目标：用一张 4×3 网格，把所有相关人脉”按位置、定策略”，持续复盘，最大化资源利用与风险管控。
Published:
我现在有一个观点，不管做啥，都很像一个在市场里投机的人。
Published:
Published:
这次 Texas ML Symposium 上最让我记住的，是一场讲怎么扩展语言模型的报告，不是怎么把模型做大。
Published:
你的下一台云服务器是活的，每三天要喂一次饭，而且不管你续不续费，它六个月后都会死。
Published:
很多人会说，十年估值十个亿美金，就是一只正常的独角兽，不慢也不快。但你真去算一下复利就会发现，这「正常」二字背后藏着一个每年估值翻一倍的怪兽。
Published:
为什么AI Agent缺乏时间意识，以及注入时间戳如何将它们从无状态自动机转变为连贯的实体。
Published:
Published:
我刚刚在想，为什么我对证券的客服还挺客气，对中介却不能客气。
Published:
「十万倍」这个说法乍听像聊天里随手甩出去的量级，但如果把它放回语义信息熵和有效训练样本这两把尺子下面，它非但不夸张，反而可能是保守的。关键是你得把「原始物理比特」和「带语义的、可被学习的数据」分开看。
Published:
Published:
我有一个想法，我们身体目前的结构是适应古代的。
Published:
当大语言模型（LLM）产生幻觉时，我们通常将其视为需要修复的工程缺陷。但如果幻觉不是缺陷，而是数学上的必然结果呢？最近一场跨学科讨论揭示了傅里叶分析中的吉布斯现象与神经网络根本局限性之间的深刻联系。
Published:
直接发钱会怎样？奥特曼的全民基本收入实验
Published:
「十万倍」这个说法乍听像聊天里随手甩出去的量级，但如果把它放回语义信息熵和有效训练样本这两把尺子下面，它非但不夸张，反而可能是保守的。关键是你得把「原始物理比特」和「带语义的、可被学习的数据」分开看。
Published:
沉没成本不应该作为决策条件，选专业也一样。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
丈夫只手把吴钩，意气高于百尺楼。一万年来谁著史，三千里外欲封侯。
Published:
新闻本身不是机会。新闻制造的摩擦才是机会。
Published:
Hah, I can’t help but laugh. It just hit me: the first time I ever used Markdown was back when I was building a poetry society. And that’s also when I first learned about Git. Looking back now, if you put all the founding members of that poetry society together, you’d almost have two unicorns.
Published:
哎，我他妈笑了。我忽然想起来，我最早用 Markdown，就是之前创建诗社的时候。知道 Git，也是在那个时候。现在回头看，整个诗社的元老凑在一起，真的快有两个独角兽了。
Published:
Many Chinese teams building embodied AI or AI applications hit the same real-world problem: the domestic models aren’t enough, and they want to call the APIs of foreign models like OpenAI and Anthropic directly. That’s why a swarm of “relay,” “recharge,” and “one OpenAI-compatible interface for you” businesses has sprung up.
Published:
很多做具身智能、做 AI 应用的中国团队，都会遇到同一个现实问题：国内的模型不够用，想直接调 OpenAI、Anthropic 这些海外模型的 API。于是市面上冒出一堆”中转”“代充”“一个 OpenAI 兼容接口给你用”的生意。
Published:
Anthropic’s most ruthless move isn’t polishing every product to perfection before release, but daring to throw out a research preview first, then rapidly getting stronger in front of everyone.
Published:
Anthropic 最狠的地方，不是每次都把产品打磨到完美才发布，而是敢先把一个 research preview 扔出来，然后当着所有人的面快速变强。
Published:
很多人会说，十年估值十个亿美金，就是一只正常的独角兽，不慢也不快。但你真去算一下复利就会发现，这「正常」二字背后藏着一个每年估值翻一倍的怪兽。
Published:
科研、股市还有创业，这些所有需要展示并能得到结果的东西，本质上都是”主业/具有护城河的本行加上热点”。所以需要做好准备,或者说不一定非要进行那种极其周全的准备，只要掌握一些方法论就可以。利用 AI 在关键时刻能够产出内容，一旦关键热点到来，马上就可以抓住。
Published:
I’ve been discussing this question with a friend lately. He wants to use AI for matching in the dating market; I think with a small sample size this is perfectly feasible, you don’t even need to build any search/ads/rec system at all. Just ask the large language model directly, toss in a few users’ profiles, and have it rank them; the whole process is very simple.
Published:
我和朋友最近在讨论这个问题。朋友说想拿 AI 来做婚恋市场的匹配，我觉得在样本量小的情况下，这完全可行——甚至根本不需要搭任何搜广推系统。你直接问大语言模型，把几个用户的简历丢进去，让它排序就好了，整个过程非常简单。
Published:
The “Alumni Real-Life Archive” documents how alumni made decisions at turning points, giving current students a wider range of real paths to consider.
Published:
欢迎加入”校友真实档案”。这里记录的不是成功经验，而是人在关键节点如何做选择。我们希望通过这些真实路径的分享，为在校学生呈现人生的多样性。
Published:
The most important thing is selling. Having technology is useless; it only matters if someone is willing to buy. I don’t think anyone has a real tech moat; that part is easy to solve. As long as you have a little technical foundation, you can handle it. What matters most is that people come and buy. Though, it might be that my own technical level is too high, and I’ve grown numb to technology.
Published:
所以最重要的是推销。有技术并没有什么卵用，只要有人要购买才有用。我觉得所有人都没有 tech 壁垒，这个很容易解决。只要稍微有一点 tech 基础，都能解决。最重要的是，有人来买。不过，也可能是我的技术太高了，我对技术已经无感。
Published:
The most terrifying thing about Silicon Valley has never been the brilliant ideas of geniuses, but that it established an entire system of mechanisms that encourage “creative betrayal.”
Published:
硅谷最可怕的从来不是那些天才的想法，而是它建立了一整套鼓励「创造性背叛」的机制。
Published:
五位 YC 叛逆者：有人被驱逐、有人被重新接纳，后来又发生了什么？
Published:
这是我见过最糟糕的创业产品。
Published:
为什么有些人看起来没你努力，却比你成功得早？因为成功首先是一个时间问题，然后才是能力问题。
Published:
你选的专业，你以为是理性决策，其实本身就是一场赌博。
Published:
Published:
Published:
Published:
闲聊历史，美国理科生比商科、法学、医学要给的 stipend 多，因为理科生穷。
Published:
建立在 19 世纪学徒制模型上的博士培养体系，几乎在所有关键经验指标上都已经出现了系统性失灵。
Published:
Doing a PhD cannot rely on “felt effort” and “moving yourself emotionally.” We need to upgrade day-to-day literature reading, experiment design, data analysis, and paper writing into a personal research system that is quantifiable, reviewable, and continuously optimizable. The point is not self-exploitation, but verifying whether your research efficiency truly exists, and keeping precious PhD time from being consumed inefficiently.
Published:
读博不能仅凭“感觉努力”和“自我感动”。我们需要将日常的文献阅读、实验设计、数据分析和论文写作，升级为一套可量化、可复盘、可持续优化的个人科研系统。重点不在于自我压榨，而在于验证自己的科研效率是否真实存在，避免宝贵的博士时间被低效消耗。
Published:
今天有个学生来咨询暑研申请，我问他目标是什么，他直接说，要拿 return offer。
Published:
在中国长大的孩子，留学申请上遇到的第一个问题，往往也是最大的问题，不是做什么事儿，而是在想什么东西。
Published:
今天我站在了圣安东尼奥的阿拉莫教堂前。这座看似宁静的建筑，却承载着德克萨斯乃至美国历史上最惨烈、最具传奇色彩的一页。
Published:
同时拿到两个暑研面试，你第一反应是开心，还是害怕把两个老师都得罪了？
Published:
我有一个想法，我们身体目前的结构是适应古代的。
Published:
我在纽约大都会艺术博物馆，走进了一座真实的古埃及墓室，不是复制品，不是模型，是四千多年前立在埃及沙漠里的墓室，整个被搬到了曼哈顿。
Published:
你选的专业，你以为是理性决策，其实本身就是一场赌博。
Published:
摆摊的人赚的是差价，做成企业的人赚的是估值。两者之间隔着一个问题：谁能先拿到别人的钱。
Published:
现在越来越觉得，只有做生意，只有商业，只有贸易，才是最根本的慈善。
Published:
我走错了温哥华机场的 NEXUS 通道；规则上是我的错，设计上却是一条几乎可以预测的错误路径。
Published:
除了那次走错 NEXUS 通道的导视问题，这次在温哥华机场（YVR）转机还让我观察到几件和路标无关、但同样值得记录的事情：明显偏大的现场年龄结构、几乎不为办公设计的候机空间，以及温哥华与得州人群景观的直观差异。这篇笔记单独记录这些观察；如果你想看我是怎么在 NEXUS 通道走错路的，可以读我在 YVR”合理地”走错了 NEXUS。
Published:
这次从新加坡经温哥华飞回奥斯汀，我对 Air Canada 的评价很简单：太拉了。
Published:
一份由 AI 制定的完美旅行计划，最后可能会死在一个不存在的“重试”按钮上。
Published:
从来没有什么最顶级的、普遍适用的方法论，只有一个，就是在一个行业里面深耕，建立个人影响力，建立个人品牌。
Published:
时隔两年，又来到这熟悉的城市。
Published:
单字母 .ai 域名只有 26 个。我把它们挨个访问了一遍，看谁在真做事，谁在蹲着等涨价。答案：大约六个在跑真产品，一个属于马斯克，一个属于谷歌，一个据说属于苹果但没有网站，剩下的是开价 150 万到 5 亿美元的出售页。以下是完整字母表。
Published:
由神经形态计算和基因语言模型驱动的虚拟细胞模型，代表了生物技术 AI 领域中资本最密集、科学最雄心勃勃的融合方向之一。如果你正在以 VC 实习生的身份评估这个领域，你需要理解三个核心组成部分：这些系统实际上做什么、为什么市场现在开始行动，以及可投资的机会在哪里。
Published:
Published:
我有一个想法，我们身体目前的结构是适应古代的。
Published:
工作中突然冒出一件紧急的事，又没有时间准备，只能熬夜，这其实就是一种黑天鹅。
Published:
很多人会说，十年估值十个亿美金，就是一只正常的独角兽，不慢也不快。但你真去算一下复利就会发现，这「正常」二字背后藏着一个每年估值翻一倍的怪兽。
Published:
我曾经以为，自己因为好玩在 LinkedIn 上手动添加几千个人，是假把式，没有任何用。
Published:
不要想第一条视频就能爆款，不要想第一次出去找生意就一定能拿到一千万，但是你可以想想，100 次之后会怎样。
Published:
想要直观的感受到自己成长，就每天写日记。
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
Compute / tokens are the new generation of free eggs. Right now all the big vendors are handing out compute like they’re handing out eggs. When you’re old, you’ll be fighting other old grannies and grandpas for free eggs, this is no longer just a joke, it might be reality.
Published:
算力 / token 是新一代的鸡蛋。现在各大厂商都在发算力，就像在发鸡蛋一样。你老了就要去跟别的老奶奶、老大爷抢发鸡蛋——这已经不是一个笑话，可能就是一个现实。
Published:
这次 Texas ML Symposium 上最让我记住的，是一场讲怎么扩展语言模型的报告，不是怎么把模型做大。
Published:
当大语言模型（LLM）产生幻觉时，我们通常将其视为需要修复的工程缺陷。但如果幻觉不是缺陷，而是数学上的必然结果呢？最近一场跨学科讨论揭示了傅里叶分析中的吉布斯现象与神经网络根本局限性之间的深刻联系。
Published:
我在纽约大都会艺术博物馆，走进了一座真实的古埃及墓室，不是复制品，不是模型，是四千多年前立在埃及沙漠里的墓室，整个被搬到了曼哈顿。
Published:
@elonmusk，替朋友问一句，特斯拉 FSD 在这种天气里测过吗？
Published:
1988 年一个周日的下午，风雨大作，图灵奖得主，卡耐基梅隆大学教授埃德蒙·克拉克来复旦讲课，听众寥寥。
Published:
直接发钱会怎样？奥特曼的全民基本收入实验
Published:
Published:
我意识到，我之前对于上课无用论，有一个非常严重的误解。
Published:
我觉得高中以后的上课 a.e.（almost everywhere，几乎处处）没有任何意义。
Published:
我越来越鄙弃和蔑视「学习」这个词。
Published:
1988 年一个周日的下午，风雨大作，图灵奖得主，卡耐基梅隆大学教授埃德蒙·克拉克来复旦讲课，听众寥寥。
Published:
When a random idea hits you on a US campus to slip a ballistic plate into the North Face backpack you carry every day, things start racing toward cyberpunk.
Published:
当你在美国校园里突发奇想，想给每天背的The North Face书包塞一块防弹板时，事情就开始朝着赛博朋克的方向狂奔了。
Published:
顶尖的 physics 老年研究者并非看不见应用层的繁荣，而是被自己的评价体系与资源锁定。
Published:
其实在从 SFO 回家这一路上，我挺伤心的。
Published:
我走错了温哥华机场的 NEXUS 通道；规则上是我的错，设计上却是一条几乎可以预测的错误路径。
Published:
前几天杨立昆（Yann LeCun）在 LinkedIn 上发帖，说自己担任了一个新的职位，这在国内是几乎不可以想象的。
Published:
多年以后，Constantin 还是忘不了那个晚上：他第一次在 Esther 的车上，听到 Somewhere Only We Know。
Published:
一杯奶茶，小费明明选的是 0，店家照样给你强制加 18%。
Published:
拍摄黑洞照片和训练GPT-4在根本上是同一个过程：从极度稀疏的频域测量中提取相干结构。
Published:
现在越来越觉得，只有做生意，只有商业，只有贸易，才是最根本的慈善。
Published:
今天下午，我在奥斯汀的动物收容所花了 80 美元，把一只三岁的橘猫领回了家。
Published:
@elonmusk，替朋友问一句，特斯拉 FSD 在这种天气里测过吗？
Published:
物理定律在任何坐标系里都写成同一个样子，这句话听着像废话，爱因斯坦为它折腾了八年，中间还亲手判过它死刑。
Published:
翁家羿在 OpenAI 的职业规划逻辑，是最大化自己名字出现在 OpenAI blog 上的次数。
Published:
In statistics, how you find outliers depends on the dimensionality of your data, its distributional characteristics, and your tolerance for what counts as “abnormal.” Here are the most common and standard approaches:
Published:
在统计学中，寻找离群值（Outlier）的方法取决于数据的维度、分布特征以及你对“异常”的容忍程度。以下是几种最常用且标准的方法：
Published:
翁家羿在 OpenAI 的职业规划逻辑，是最大化自己名字出现在 OpenAI blog 上的次数。
Published:
微软和 OpenAI 在解放人类这个终极目标上高度同构，区别只在杠杆撬到了哪一层。
Published:
今天我站在了圣安东尼奥的阿拉莫教堂前。这座看似宁静的建筑，却承载着德克萨斯乃至美国历史上最惨烈、最具传奇色彩的一页。
Published:
同时拿到两个暑研面试，你第一反应是开心，还是害怕把两个老师都得罪了？
Published:
在中国长大的孩子，留学申请上遇到的第一个问题，往往也是最大的问题，不是做什么事儿，而是在想什么东西。
Published:
When a random idea hits you on a US campus to slip a ballistic plate into the North Face backpack you carry every day, things start racing toward cyberpunk.
Published:
当你在美国校园里突发奇想，想给每天背的The North Face书包塞一块防弹板时，事情就开始朝着赛博朋克的方向狂奔了。
Published:
很多人会说，十年估值十个亿美金，就是一只正常的独角兽，不慢也不快。但你真去算一下复利就会发现，这「正常」二字背后藏着一个每年估值翻一倍的怪兽。
Published:
顶尖的 physics 老年研究者并非看不见应用层的繁荣，而是被自己的评价体系与资源锁定。
Published:
The most important thing is selling. Having technology is useless; it only matters if someone is willing to buy. I don’t think anyone has a real tech moat; that part is easy to solve. As long as you have a little technical foundation, you can handle it. What matters most is that people come and buy. Though, it might be that my own technical level is too high, and I’ve grown numb to technology.
Published:
所以最重要的是推销。有技术并没有什么卵用，只要有人要购买才有用。我觉得所有人都没有 tech 壁垒，这个很容易解决。只要稍微有一点 tech 基础，都能解决。最重要的是，有人来买。不过，也可能是我的技术太高了，我对技术已经无感。
Published:
The most terrifying thing about Silicon Valley has never been the brilliant ideas of geniuses, but that it established an entire system of mechanisms that encourage “creative betrayal.”
Published:
硅谷最可怕的从来不是那些天才的想法，而是它建立了一整套鼓励「创造性背叛」的机制。
Published:
我花了一个晚上跟 Gemini 把美国富豪那套 Buy Borrow Die 的玩法从头扒到尾，本来以为只要照抄就能复刻，结果一层层剥下来才发现，F-1 学生在这个剧本里几乎处处是坑，能用的招其实只有那么三两个。
Published:
金涨多了就换成银，银涨多了再换回金，这个听起来像民间偏方的操作，其实是贵金属市场里最经典的增厚策略之一。
Published:
微软和 OpenAI 在解放人类这个终极目标上高度同构，区别只在杠杆撬到了哪一层。
Published:
现在越来越觉得，只有做生意，只有商业，只有贸易，才是最根本的慈善。
Published:
想要直观的感受到自己成长，就每天写日记。
Published:
The “Alumni Real-Life Archive” documents how alumni made decisions at turning points, giving current students a wider range of real paths to consider.
Published:
欢迎加入”校友真实档案”。这里记录的不是成功经验，而是人在关键节点如何做选择。我们希望通过这些真实路径的分享，为在校学生呈现人生的多样性。
Published:
一份由 AI 制定的完美旅行计划，最后可能会死在一个不存在的“重试”按钮上。
Published:
我越来越鄙弃和蔑视「学习」这个词。
Published:
The most important thing is selling. Having technology is useless; it only matters if someone is willing to buy. I don’t think anyone has a real tech moat; that part is easy to solve. As long as you have a little technical foundation, you can handle it. What matters most is that people come and buy. Though, it might be that my own technical level is too high, and I’ve grown numb to technology.
Published:
所以最重要的是推销。有技术并没有什么卵用，只要有人要购买才有用。我觉得所有人都没有 tech 壁垒，这个很容易解决。只要稍微有一点 tech 基础，都能解决。最重要的是，有人来买。不过，也可能是我的技术太高了，我对技术已经无感。
Published:
我花了一个晚上跟 Gemini 把美国富豪那套 Buy Borrow Die 的玩法从头扒到尾，本来以为只要照抄就能复刻，结果一层层剥下来才发现，F-1 学生在这个剧本里几乎处处是坑，能用的招其实只有那么三两个。
Published:
金涨多了就换成银，银涨多了再换回金，这个听起来像民间偏方的操作，其实是贵金属市场里最经典的增厚策略之一。
Published:
今天又是高强度网上冲浪的一天，学到最多的一点是，世界上没有躺着就能财富自由的投资策略。
Published:
发紧急邮件求教的时候，用 quick question 比写 emergency 更容易收到回复。
Published:
其实在从 SFO 回家这一路上，我挺伤心的。
Published:
I’ve been discussing this question with a friend lately. He wants to use AI for matching in the dating market; I think with a small sample size this is perfectly feasible, you don’t even need to build any search/ads/rec system at all. Just ask the large language model directly, toss in a few users’ profiles, and have it rank them; the whole process is very simple.
Published:
我和朋友最近在讨论这个问题。朋友说想拿 AI 来做婚恋市场的匹配，我觉得在样本量小的情况下，这完全可行——甚至根本不需要搭任何搜广推系统。你直接问大语言模型，把几个用户的简历丢进去，让它排序就好了，整个过程非常简单。
Published:
翁家羿在 OpenAI 的职业规划逻辑，是最大化自己名字出现在 OpenAI blog 上的次数。
Published:
一份由 AI 制定的完美旅行计划，最后可能会死在一个不存在的“重试”按钮上。
Published:
目标：用一张 4×3 网格，把所有相关人脉”按位置、定策略”，持续复盘，最大化资源利用与风险管控。
Published:
Doing a PhD cannot rely on “felt effort” and “moving yourself emotionally.” We need to upgrade day-to-day literature reading, experiment design, data analysis, and paper writing into a personal research system that is quantifiable, reviewable, and continuously optimizable. The point is not self-exploitation, but verifying whether your research efficiency truly exists, and keeping precious PhD time from being consumed inefficiently.
Published:
读博不能仅凭“感觉努力”和“自我感动”。我们需要将日常的文献阅读、实验设计、数据分析和论文写作，升级为一套可量化、可复盘、可持续优化的个人科研系统。重点不在于自我压榨，而在于验证自己的科研效率是否真实存在，避免宝贵的博士时间被低效消耗。
Published:
不要想第一条视频就能爆款，不要想第一次出去找生意就一定能拿到一千万，但是你可以想想，100 次之后会怎样。
Published:
On a United Airlines flight from San Francisco SFO to Shanghai Pudong PVG, the plane had started its descent and the cabin announcements kept reminding everyone to fasten their seatbelts. Outside was night; inside, people were already lit up with the excitement of “finally going home.” It was on this flight that I met a friend.
Published:
在联合航空从旧金山 SFO 飞往上海浦东 PVG 的航班上，飞机已经开始下降，机舱广播一遍遍提醒大家系好安全带。窗外是夜色，舱内的人却已经被“终于回国了”的兴奋点亮。我就在这趟航班上认识了一位朋友。
Published:
我用了二十多年才培育起来的主体性，这些十七八岁的留子，就因为在 Stanford，在美国，或许更早在上海读了高中，就轻易许多地拿到了。
Published:
我意识到，我之前对于上课无用论，有一个非常严重的误解。
Published:
今天回看之前的日记，从物理的核聚变，到人类的把戏，无论一个 startup 还是一个行业，有一点很重要，就是自发涌动乃至永动的力量。
Published:
我觉得高中以后的上课 a.e.（almost everywhere，几乎处处）没有任何意义。
Published:
2024 年，在 40 余位学者中，我是唯一受邀的本科生。
Published:
为什么Next Token Prediction不只是”随机鹦鹉”——它在求解人类知识流形上的测地线。
Published:
当连续函数遇上离散真理：从傅里叶分析的视角理解LLM幻觉。
Published:
当大语言模型（LLM）产生幻觉时，我们通常将其视为需要修复的工程缺陷。但如果幻觉不是缺陷，而是数学上的必然结果呢？最近一场跨学科讨论揭示了傅里叶分析中的吉布斯现象与神经网络根本局限性之间的深刻联系。
Published:
In statistics, how you find outliers depends on the dimensionality of your data, its distributional characteristics, and your tolerance for what counts as “abnormal.” Here are the most common and standard approaches:
Published:
在统计学中，寻找离群值（Outlier）的方法取决于数据的维度、分布特征以及你对“异常”的容忍程度。以下是几种最常用且标准的方法：
Published:
今天我站在了圣安东尼奥的阿拉莫教堂前。这座看似宁静的建筑，却承载着德克萨斯乃至美国历史上最惨烈、最具传奇色彩的一页。
Published:
Although “umami” is the unshakable soul of Jiangsu-Zhejiang and Cantonese cooking, in traditional American food you often wander only among single-note salt, sweet, and oil, struggling to find that layered depth of flavor. That isn’t accidental; it’s the inevitable result of a deep cultural difference in how ingredients are handled, how seasoning is reasoned, and how industrial production works.
Published:
虽然“鲜”（Umami）在江浙菜和广府菜中是不可撼动的灵魂，但在传统美国菜里，你往往只能在单一的咸、甜、油之间徘徊，而难觅那种富有层次感的味觉深度。这并非偶然，而是一场关于食材处理、调味逻辑与工业化生产方式的深层文化差异所导致的必然结果。
Published:
为什么有些人看起来没你努力，却比你成功得早？因为成功首先是一个时间问题，然后才是能力问题。
Published:
其实在从 SFO 回家这一路上，我挺伤心的。
Published:
目标：用一张 4×3 网格，把所有相关人脉”按位置、定策略”，持续复盘，最大化资源利用与风险管控。
Published:
On a United Airlines flight from San Francisco SFO to Shanghai Pudong PVG, the plane had started its descent and the cabin announcements kept reminding everyone to fasten their seatbelts. Outside was night; inside, people were already lit up with the excitement of “finally going home.” It was on this flight that I met a friend.
Published:
在联合航空从旧金山 SFO 飞往上海浦东 PVG 的航班上，飞机已经开始下降，机舱广播一遍遍提醒大家系好安全带。窗外是夜色，舱内的人却已经被“终于回国了”的兴奋点亮。我就在这趟航班上认识了一位朋友。
Published:
今天我站在了圣安东尼奥的阿拉莫教堂前。这座看似宁静的建筑，却承载着德克萨斯乃至美国历史上最惨烈、最具传奇色彩的一页。
Published:
我最近从一位去过日本多次的朋友那里得到一句很直接的建议：去日本之前，先把行程排明白。
Published:
这次从新加坡经温哥华飞回奥斯汀，我对 Air Canada 的评价很简单：太拉了。
Published:
我最近从一位去过日本多次的朋友那里得到一句很直接的建议：去日本之前，先把行程排明白。
Published:
想要直观的感受到自己成长，就每天写日记。
Published:
Published:
为什么AI Agent缺乏时间意识，以及注入时间戳如何将它们从无状态自动机转变为连贯的实体。
Published:
今天有个学生来咨询暑研申请，我问他目标是什么，他直接说，要拿 return offer。
Published:
多年以后，Constantin 还是忘不了那个晚上：他第一次在 Esther 的车上，听到 Somewhere Only We Know。
Published:
Published:
同时拿到两个暑研面试，你第一反应是开心，还是害怕把两个老师都得罪了？
Published:
今天有个学生来咨询暑研申请，我问他目标是什么，他直接说，要拿 return offer。
Published:
我刚刚在想，为什么我对证券的客服还挺客气，对中介却不能客气。
Published:
直接发钱会怎样？奥特曼的全民基本收入实验
Published:
Published:
「十万倍」这个说法乍听像聊天里随手甩出去的量级，但如果把它放回语义信息熵和有效训练样本这两把尺子下面，它非但不夸张，反而可能是保守的。关键是你得把「原始物理比特」和「带语义的、可被学习的数据」分开看。
Published:
为什么Next Token Prediction不只是”随机鹦鹉”——它在求解人类知识流形上的测地线。
Published:
当连续函数遇上离散真理：从傅里叶分析的视角理解LLM幻觉。
Published:
拍摄黑洞照片和训练GPT-4在根本上是同一个过程：从极度稀疏的频域测量中提取相干结构。
Published:
这次 Texas ML Symposium 上最让我记住的，是一场讲怎么扩展语言模型的报告，不是怎么把模型做大。
Published:
我走错了温哥华机场的 NEXUS 通道；规则上是我的错，设计上却是一条几乎可以预测的错误路径。
Published:
除了那次走错 NEXUS 通道的导视问题，这次在温哥华机场（YVR）转机还让我观察到几件和路标无关、但同样值得记录的事情：明显偏大的现场年龄结构、几乎不为办公设计的候机空间，以及温哥华与得州人群景观的直观差异。这篇笔记单独记录这些观察；如果你想看我是怎么在 NEXUS 通道走错路的，可以读我在 YVR”合理地”走错了 NEXUS。
Published:
Published:
Published:
Published:
微软和 OpenAI 在解放人类这个终极目标上高度同构，区别只在杠杆撬到了哪一层。
Published:
Published:
Published:
一张 1957 年的诺贝尔奖支票，一个中国第一位数论博士的父亲，三套不同的货币制度，加上七十年的复利，我们到底要怎么给杨振宁家族 2026 年的财富估出一个不是瞎编的数字？
Published:
The “Alumni Real-Life Archive” documents how alumni made decisions at turning points, giving current students a wider range of real paths to consider.
Published:
欢迎加入”校友真实档案”。这里记录的不是成功经验，而是人在关键节点如何做选择。我们希望通过这些真实路径的分享，为在校学生呈现人生的多样性。
Published:
| Original source | Tencent Tech; transcript edited | Gu Lingyu; original editors | Xu Qingyang, Su Yang; 2030 annotations and second edit | Koutian Wu |
Published:
原文来源｜腾讯科技；原文整理｜顾翎羽；原文编辑｜徐青阳、苏扬；2030 注释与二次编辑｜Koutian Wu
Published:
摆摊的人赚的是差价，做成企业的人赚的是估值。两者之间隔着一个问题：谁能先拿到别人的钱。
Published:
Published:
沉没成本不应该作为决策条件，选专业也一样。
Published:
发紧急邮件求教的时候，用 quick question 比写 emergency 更容易收到回复。
Published:
为什么Next Token Prediction不只是”随机鹦鹉”——它在求解人类知识流形上的测地线。
Published:
我在纽约大都会艺术博物馆，走进了一座真实的古埃及墓室，不是复制品，不是模型，是四千多年前立在埃及沙漠里的墓室，整个被搬到了曼哈顿。
Published:
从library出来，有位头发花白的白人老奶奶看到我们，说了三遍excuse me。她非常proudly和我们讲话，掏出手机，说，看到你们俩在和这里的青蛙雕塑拍照。每逢圣诞节，这个青蛙都会穿上圣诞套装噢~瞧，就是这样！然后很真诚地问我们会不会在这里待到圣诞，问我们怎么知道Atherton这个地方的，因为这里实在是太不起眼、太静谧、太低调了。我俩说，可能要各回各家了。老奶奶又说她儿子从小不点的时候开始就学中文了，说的可六。
Published:
来美国读了 PhD 一年半，我得到最好的训练是学会讲好故事。
Published:
2024 年，在 40 余位学者中，我是唯一受邀的本科生。
Published:
物理定律在任何坐标系里都写成同一个样子，这句话听着像废话，爱因斯坦为它折腾了八年，中间还亲手判过它死刑。
Published:
物理定律在任何坐标系里都写成同一个样子，这句话听着像废话，爱因斯坦为它折腾了八年，中间还亲手判过它死刑。
Published:
顶尖的 physics 老年研究者并非看不见应用层的繁荣，而是被自己的评价体系与资源锁定。
Published:
拍摄黑洞照片和训练GPT-4在根本上是同一个过程：从极度稀疏的频域测量中提取相干结构。
Published:
@elonmusk，替朋友问一句，特斯拉 FSD 在这种天气里测过吗？
Published:
很多人会说，十年估值十个亿美金，就是一只正常的独角兽，不慢也不快。但你真去算一下复利就会发现，这「正常」二字背后藏着一个每年估值翻一倍的怪兽。
Published:
今天下午，我在奥斯汀的动物收容所花了 80 美元，把一只三岁的橘猫领回了家。
Published:
闲聊历史，美国理科生比商科、法学、医学要给的 stipend 多，因为理科生穷。
Published:
从library出来，有位头发花白的白人老奶奶看到我们，说了三遍excuse me。她非常proudly和我们讲话，掏出手机，说，看到你们俩在和这里的青蛙雕塑拍照。每逢圣诞节，这个青蛙都会穿上圣诞套装噢~瞧，就是这样！然后很真诚地问我们会不会在这里待到圣诞，问我们怎么知道Atherton这个地方的，因为这里实在是太不起眼、太静谧、太低调了。我俩说，可能要各回各家了。老奶奶又说她儿子从小不点的时候开始就学中文了，说的可六。
Published:
由神经形态计算和基因语言模型驱动的虚拟细胞模型，代表了生物技术 AI 领域中资本最密集、科学最雄心勃勃的融合方向之一。如果你正在以 VC 实习生的身份评估这个领域，你需要理解三个核心组成部分：这些系统实际上做什么、为什么市场现在开始行动，以及可投资的机会在哪里。
Published:
你的下一台云服务器是活的，每三天要喂一次饭，而且不管你续不续费，它六个月后都会死。
Published:
我用了二十多年才培育起来的主体性，这些十七八岁的留子，就因为在 Stanford，在美国，或许更早在上海读了高中，就轻易许多地拿到了。
Published:
今天在 Palo Alto 的 University Avenue 走了一圈，我反而想通了一件事，我们很多中国留学生，把硅谷当成了一个目的地。
Published:
1988 年一个周日的下午，风雨大作，图灵奖得主，卡耐基梅隆大学教授埃德蒙·克拉克来复旦讲课，听众寥寥。
Published:
同时拿到两个暑研面试，你第一反应是开心，还是害怕把两个老师都得罪了？
Published:
今天有个学生来咨询暑研申请，我问他目标是什么，他直接说，要拿 return offer。
Published:
沉没成本不应该作为决策条件，选专业也一样。
Published:
为什么不直接搞 2000 万，3000 万？
Published:
一杯奶茶，小费明明选的是 0，店家照样给你强制加 18%。
Published:
你选的专业，你以为是理性决策，其实本身就是一场赌博。
Published:
在中国长大的孩子，留学申请上遇到的第一个问题，往往也是最大的问题，不是做什么事儿，而是在想什么东西。
Published:
我花了一个晚上跟 Gemini 把美国富豪那套 Buy Borrow Die 的玩法从头扒到尾，本来以为只要照抄就能复刻，结果一层层剥下来才发现，F-1 学生在这个剧本里几乎处处是坑，能用的招其实只有那么三两个。
Published:
从来没有什么最顶级的、普遍适用的方法论，只有一个，就是在一个行业里面深耕，建立个人影响力，建立个人品牌。
Published:
来美国读了 PhD 一年半，我得到最好的训练是学会讲好故事。
Published:
我有一个想法，我们身体目前的结构是适应古代的。
Published:
我现在有一个观点，不管做啥，都很像一个在市场里投机的人。
Published:
在学校里，考试就那么几次，你根本没有多少机会练。
Published:
金涨多了就换成银，银涨多了再换回金，这个听起来像民间偏方的操作，其实是贵金属市场里最经典的增厚策略之一。
Published:
为什么不直接搞 2000 万，3000 万？
Published:
我最近发现一件很丢人的事，AI 额度还没用完，我就睡不着觉。
Published:
一篇新论文把 1901 到 2023 年所有 739 位科学类诺贝尔奖得主从童年扒到现在，得出了一个特别狠的数字，按目前的进步速度，一个出生在低收入国家的孩子要等大约 600 年，才能拿到和富裕国家孩子一样的诺奖机会。
Published:
顶尖的 physics 老年研究者并非看不见应用层的繁荣，而是被自己的评价体系与资源锁定。
Published:
建立在 19 世纪学徒制模型上的博士培养体系，几乎在所有关键经验指标上都已经出现了系统性失灵。
Published:
Doing a PhD cannot rely on “felt effort” and “moving yourself emotionally.” We need to upgrade day-to-day literature reading, experiment design, data analysis, and paper writing into a personal research system that is quantifiable, reviewable, and continuously optimizable. The point is not self-exploitation, but verifying whether your research efficiency truly exists, and keeping precious PhD time from being consumed inefficiently.
Published:
读博不能仅凭“感觉努力”和“自我感动”。我们需要将日常的文献阅读、实验设计、数据分析和论文写作，升级为一套可量化、可复盘、可持续优化的个人科研系统。重点不在于自我压榨，而在于验证自己的科研效率是否真实存在，避免宝贵的博士时间被低效消耗。
Published:
一张 1957 年的诺贝尔奖支票，一个中国第一位数论博士的父亲，三套不同的货币制度，加上七十年的复利，我们到底要怎么给杨振宁家族 2026 年的财富估出一个不是瞎编的数字？
Published:
The most terrifying thing about Silicon Valley has never been the brilliant ideas of geniuses, but that it established an entire system of mechanisms that encourage “creative betrayal.”
Published:
硅谷最可怕的从来不是那些天才的想法，而是它建立了一整套鼓励「创造性背叛」的机制。
Published:
今天在 Palo Alto 的 University Avenue 走了一圈，我反而想通了一件事，我们很多中国留学生，把硅谷当成了一个目的地。
Published:
前几天杨立昆（Yann LeCun）在 LinkedIn 上发帖，说自己担任了一个新的职位，这在国内是几乎不可以想象的。
Published:
一张 1957 年的诺贝尔奖支票，一个中国第一位数论博士的父亲，三套不同的货币制度，加上七十年的复利，我们到底要怎么给杨振宁家族 2026 年的财富估出一个不是瞎编的数字？
Published:
在学校里，考试就那么几次，你根本没有多少机会练。
Published:
我走错了温哥华机场的 NEXUS 通道；规则上是我的错，设计上却是一条几乎可以预测的错误路径。
Published:
除了那次走错 NEXUS 通道的导视问题，这次在温哥华机场（YVR）转机还让我观察到几件和路标无关、但同样值得记录的事情：明显偏大的现场年龄结构、几乎不为办公设计的候机空间，以及温哥华与得州人群景观的直观差异。这篇笔记单独记录这些观察；如果你想看我是怎么在 NEXUS 通道走错路的，可以读我在 YVR”合理地”走错了 NEXUS。
Published:
今天回看之前的日记，从物理的核聚变，到人类的把戏，无论一个 startup 还是一个行业，有一点很重要，就是自发涌动乃至永动的力量。
Published:
由神经形态计算和基因语言模型驱动的虚拟细胞模型，代表了生物技术 AI 领域中资本最密集、科学最雄心勃勃的融合方向之一。如果你正在以 VC 实习生的身份评估这个领域，你需要理解三个核心组成部分：这些系统实际上做什么、为什么市场现在开始行动，以及可投资的机会在哪里。
Published:
一张 1957 年的诺贝尔奖支票，一个中国第一位数论博士的父亲，三套不同的货币制度，加上七十年的复利，我们到底要怎么给杨振宁家族 2026 年的财富估出一个不是瞎编的数字？
Published:
一篇新论文把 1901 到 2023 年所有 739 位科学类诺贝尔奖得主从童年扒到现在，得出了一个特别狠的数字，按目前的进步速度，一个出生在低收入国家的孩子要等大约 600 年，才能拿到和富裕国家孩子一样的诺奖机会。
Published:
物理定律在任何坐标系里都写成同一个样子，这句话听着像废话，爱因斯坦为它折腾了八年，中间还亲手判过它死刑。
Published:
科研、股市还有创业，这些所有需要展示并能得到结果的东西，本质上都是”主业/具有护城河的本行加上热点”。所以需要做好准备,或者说不一定非要进行那种极其周全的准备，只要掌握一些方法论就可以。利用 AI 在关键时刻能够产出内容，一旦关键热点到来，马上就可以抓住。
Published:
The “Alumni Real-Life Archive” documents how alumni made decisions at turning points, giving current students a wider range of real paths to consider.
Published:
欢迎加入”校友真实档案”。这里记录的不是成功经验，而是人在关键节点如何做选择。我们希望通过这些真实路径的分享，为在校学生呈现人生的多样性。
Published:
搞科研就是赌博。事实上，没钱的人搞科研是赌博，搞什么也是赌博。
Published:
You apply to Anthropic, land an offer, then ask HR whether they’ll sponsor your PERM, and they say “on hold for now” or change the subject, and you probably stand there frozen.
Published:
你投了 Anthropic，拿到了 offer，然后去问 HR 能不能帮你办 PERM，对方说「目前暂停」或者直接转移话题，你当时大概就愣在原地了。
Published:
我花了一个晚上跟 Gemini 把美国富豪那套 Buy Borrow Die 的玩法从头扒到尾，本来以为只要照抄就能复刻，结果一层层剥下来才发现，F-1 学生在这个剧本里几乎处处是坑，能用的招其实只有那么三两个。
Published:
「十万倍」这个说法乍听像聊天里随手甩出去的量级，但如果把它放回语义信息熵和有效训练样本这两把尺子下面，它非但不夸张，反而可能是保守的。关键是你得把「原始物理比特」和「带语义的、可被学习的数据」分开看。
Published:
科研、股市还有创业，这些所有需要展示并能得到结果的东西，本质上都是”主业/具有护城河的本行加上热点”。所以需要做好准备,或者说不一定非要进行那种极其周全的准备，只要掌握一些方法论就可以。利用 AI 在关键时刻能够产出内容，一旦关键热点到来，马上就可以抓住。
Published:
为什么不直接搞 2000 万，3000 万？
Published:
You apply to Anthropic, land an offer, then ask HR whether they’ll sponsor your PERM, and they say “on hold for now” or change the subject, and you probably stand there frozen.
Published:
你投了 Anthropic，拿到了 offer，然后去问 HR 能不能帮你办 PERM，对方说「目前暂停」或者直接转移话题，你当时大概就愣在原地了。
Published:
Compute / tokens are the new generation of free eggs. Right now all the big vendors are handing out compute like they’re handing out eggs. When you’re old, you’ll be fighting other old grannies and grandpas for free eggs, this is no longer just a joke, it might be reality.
Published:
算力 / token 是新一代的鸡蛋。现在各大厂商都在发算力，就像在发鸡蛋一样。你老了就要去跟别的老奶奶、老大爷抢发鸡蛋——这已经不是一个笑话，可能就是一个现实。
Published:
工作中突然冒出一件紧急的事，又没有时间准备，只能熬夜，这其实就是一种黑天鹅。
Published:
Published:
我在纽约大都会艺术博物馆，走进了一座真实的古埃及墓室，不是复制品，不是模型，是四千多年前立在埃及沙漠里的墓室，整个被搬到了曼哈顿。
Published:
In statistics, how you find outliers depends on the dimensionality of your data, its distributional characteristics, and your tolerance for what counts as “abnormal.” Here are the most common and standard approaches:
Published:
在统计学中，寻找离群值（Outlier）的方法取决于数据的维度、分布特征以及你对“异常”的容忍程度。以下是几种最常用且标准的方法：
Published:
You apply to Anthropic, land an offer, then ask HR whether they’ll sponsor your PERM, and they say “on hold for now” or change the subject, and you probably stand there frozen.
Published:
你投了 Anthropic，拿到了 offer，然后去问 HR 能不能帮你办 PERM，对方说「目前暂停」或者直接转移话题，你当时大概就愣在原地了。
Published:
摆摊的人赚的是差价，做成企业的人赚的是估值。两者之间隔着一个问题：谁能先拿到别人的钱。
Published:
On a United Airlines flight from San Francisco SFO to Shanghai Pudong PVG, the plane had started its descent and the cabin announcements kept reminding everyone to fasten their seatbelts. Outside was night; inside, people were already lit up with the excitement of “finally going home.” It was on this flight that I met a friend.
Published:
在联合航空从旧金山 SFO 飞往上海浦东 PVG 的航班上，飞机已经开始下降，机舱广播一遍遍提醒大家系好安全带。窗外是夜色，舱内的人却已经被“终于回国了”的兴奋点亮。我就在这趟航班上认识了一位朋友。
Published:
一杯奶茶，小费明明选的是 0，店家照样给你强制加 18%。
Published:
Although “umami” is the unshakable soul of Jiangsu-Zhejiang and Cantonese cooking, in traditional American food you often wander only among single-note salt, sweet, and oil, struggling to find that layered depth of flavor. That isn’t accidental; it’s the inevitable result of a deep cultural difference in how ingredients are handled, how seasoning is reasoned, and how industrial production works.
Published:
虽然“鲜”（Umami）在江浙菜和广府菜中是不可撼动的灵魂，但在传统美国菜里，你往往只能在单一的咸、甜、油之间徘徊，而难觅那种富有层次感的味觉深度。这并非偶然，而是一场关于食材处理、调味逻辑与工业化生产方式的深层文化差异所导致的必然结果。
Published:
顶尖的 physics 老年研究者并非看不见应用层的繁荣，而是被自己的评价体系与资源锁定。
Published:
科研、股市还有创业，这些所有需要展示并能得到结果的东西，本质上都是”主业/具有护城河的本行加上热点”。所以需要做好准备,或者说不一定非要进行那种极其周全的准备，只要掌握一些方法论就可以。利用 AI 在关键时刻能够产出内容，一旦关键热点到来，马上就可以抓住。
Published:
从来没有什么最顶级的、普遍适用的方法论，只有一个，就是在一个行业里面深耕，建立个人影响力，建立个人品牌。
Published:
You apply to Anthropic, land an offer, then ask HR whether they’ll sponsor your PERM, and they say “on hold for now” or change the subject, and you probably stand there frozen.
Published:
你投了 Anthropic，拿到了 offer，然后去问 HR 能不能帮你办 PERM，对方说「目前暂停」或者直接转移话题，你当时大概就愣在原地了。
Published:
今天在 Palo Alto 的 University Avenue 走了一圈，我反而想通了一件事，我们很多中国留学生，把硅谷当成了一个目的地。
Published:
翁家羿在 OpenAI 的职业规划逻辑，是最大化自己名字出现在 OpenAI blog 上的次数。
Published:
搞科研就是赌博。事实上，没钱的人搞科研是赌博，搞什么也是赌博。
Published:
我现在有一个观点，不管做啥，都很像一个在市场里投机的人。
Published:
沉没成本不应该作为决策条件，选专业也一样。
Published:
为什么有些人看起来没你努力，却比你成功得早？因为成功首先是一个时间问题，然后才是能力问题。
Published:
你选的专业，你以为是理性决策，其实本身就是一场赌博。
Published:
@elonmusk，替朋友问一句，特斯拉 FSD 在这种天气里测过吗？
Published:
我意识到，我之前对于上课无用论，有一个非常严重的误解。
Published:
我最近从一位去过日本多次的朋友那里得到一句很直接的建议：去日本之前，先把行程排明白。
Published:
时隔两年，又来到这熟悉的城市。
Published:
这次从新加坡经温哥华飞回奥斯汀，我对 Air Canada 的评价很简单：太拉了。
Published:
Published:
由神经形态计算和基因语言模型驱动的虚拟细胞模型，代表了生物技术 AI 领域中资本最密集、科学最雄心勃勃的融合方向之一。如果你正在以 VC 实习生的身份评估这个领域，你需要理解三个核心组成部分：这些系统实际上做什么、为什么市场现在开始行动，以及可投资的机会在哪里。
Published:
我最近从一位去过日本多次的朋友那里得到一句很直接的建议：去日本之前，先把行程排明白。
Published:
Compute / tokens are the new generation of free eggs. Right now all the big vendors are handing out compute like they’re handing out eggs. When you’re old, you’ll be fighting other old grannies and grandpas for free eggs, this is no longer just a joke, it might be reality.
Published:
算力 / token 是新一代的鸡蛋。现在各大厂商都在发算力，就像在发鸡蛋一样。你老了就要去跟别的老奶奶、老大爷抢发鸡蛋——这已经不是一个笑话，可能就是一个现实。
Published:
When a random idea hits you on a US campus to slip a ballistic plate into the North Face backpack you carry every day, things start racing toward cyberpunk.
Published:
当你在美国校园里突发奇想，想给每天背的The North Face书包塞一块防弹板时，事情就开始朝着赛博朋克的方向狂奔了。
Published:
这题问到刀尖上了。如果把「高质量多模态训练数据」收窄到对 GPT-4V、Sora、Gemini、Emu 这类模型真有喂饭价值的图文或视频，短答其实很锋利，Meta（Facebook / Instagram）静态图片数量的库存对其他两家几乎是断层第一，ByteDance（TikTok / 抖音）在短视频也就是动态图片流上占最大优势，X（Twitter）绝对量级最小，但图文的贴脸相关性和实时信息密度是独一份。
Published:
1988 年一个周日的下午，风雨大作，图灵奖得主，卡耐基梅隆大学教授埃德蒙·克拉克来复旦讲课，听众寥寥。
Published:
来美国读了 PhD 一年半，我得到最好的训练是学会讲好故事。
Published:
在学校里，考试就那么几次，你根本没有多少机会练。
Published:
Published:
一张 1957 年的诺贝尔奖支票，一个中国第一位数论博士的父亲，三套不同的货币制度，加上七十年的复利，我们到底要怎么给杨振宁家族 2026 年的财富估出一个不是瞎编的数字？
Published:
一篇新论文把 1901 到 2023 年所有 739 位科学类诺贝尔奖得主从童年扒到现在，得出了一个特别狠的数字，按目前的进步速度，一个出生在低收入国家的孩子要等大约 600 年，才能拿到和富裕国家孩子一样的诺奖机会。
Published:
很多人会说，十年估值十个亿美金，就是一只正常的独角兽，不慢也不快。但你真去算一下复利就会发现，这「正常」二字背后藏着一个每年估值翻一倍的怪兽。
Published:
我花了一个晚上跟 Gemini 把美国富豪那套 Buy Borrow Die 的玩法从头扒到尾，本来以为只要照抄就能复刻，结果一层层剥下来才发现，F-1 学生在这个剧本里几乎处处是坑，能用的招其实只有那么三两个。
Published:
一张 1957 年的诺贝尔奖支票，一个中国第一位数论博士的父亲，三套不同的货币制度，加上七十年的复利，我们到底要怎么给杨振宁家族 2026 年的财富估出一个不是瞎编的数字？
Published:
一篇新论文把 1901 到 2023 年所有 739 位科学类诺贝尔奖得主从童年扒到现在，得出了一个特别狠的数字，按目前的进步速度，一个出生在低收入国家的孩子要等大约 600 年，才能拿到和富裕国家孩子一样的诺奖机会。
Published:
今天又是高强度网上冲浪的一天，学到最多的一点是，世界上没有躺着就能财富自由的投资策略。
Published:
金涨多了就换成银，银涨多了再换回金，这个听起来像民间偏方的操作，其实是贵金属市场里最经典的增厚策略之一。
Published:
新闻本身不是机会。新闻制造的摩擦才是机会。
Published:
我现在有一个观点，不管做啥，都很像一个在市场里投机的人。
Published:
发紧急邮件求教的时候，用 quick question 比写 emergency 更容易收到回复。
Published:
这是我见过最糟糕的创业产品。
Published:
今天又是高强度网上冲浪的一天，学到最多的一点是，世界上没有躺着就能财富自由的投资策略。
Published:
摆摊的人赚的是差价，做成企业的人赚的是估值。两者之间隔着一个问题：谁能先拿到别人的钱。
Published:
The most important thing is selling. Having technology is useless; it only matters if someone is willing to buy. I don’t think anyone has a real tech moat; that part is easy to solve. As long as you have a little technical foundation, you can handle it. What matters most is that people come and buy. Though, it might be that my own technical level is too high, and I’ve grown numb to technology.
Published:
所以最重要的是推销。有技术并没有什么卵用，只要有人要购买才有用。我觉得所有人都没有 tech 壁垒，这个很容易解决。只要稍微有一点 tech 基础，都能解决。最重要的是，有人来买。不过，也可能是我的技术太高了，我对技术已经无感。
Published:
不要想第一条视频就能爆款，不要想第一次出去找生意就一定能拿到一千万，但是你可以想想，100 次之后会怎样。
Published:
Published:
多年以后，Constantin 还是忘不了那个晚上：他第一次在 Esther 的车上，听到 Somewhere Only We Know。
Published:
Published:
Published:
今天下午，我在奥斯汀的动物收容所花了 80 美元，把一只三岁的橘猫领回了家。
Published:
There are only 26 single-letter .ai domains. I went through all of them, one by one, to see who actually builds something and who just squats. The answer: about six run real products, one belongs to Elon Musk, one belongs to Google, one apparently belongs to Apple without a website, and the rest are for-sale pages asking between 1.5 and 500 million dollars. Here is the full alphabet.
Published:
单字母 .ai 域名只有 26 个。我把它们挨个访问了一遍，看谁在真做事，谁在蹲着等涨价。答案：大约六个在跑真产品，一个属于马斯克，一个属于谷歌，一个据说属于苹果但没有网站，剩下的是开价 150 万到 5 亿美元的出售页。以下是完整字母表。
Published:
Compute / tokens are the new generation of free eggs. Right now all the big vendors are handing out compute like they’re handing out eggs. When you’re old, you’ll be fighting other old grannies and grandpas for free eggs, this is no longer just a joke, it might be reality.
Published:
算力 / token 是新一代的鸡蛋。现在各大厂商都在发算力，就像在发鸡蛋一样。你老了就要去跟别的老奶奶、老大爷抢发鸡蛋——这已经不是一个笑话，可能就是一个现实。
Published:
时隔两年，又来到这熟悉的城市。
Published:
When a random idea hits you on a US campus to slip a ballistic plate into the North Face backpack you carry every day, things start racing toward cyberpunk.
Published:
当你在美国校园里突发奇想，想给每天背的The North Face书包塞一块防弹板时，事情就开始朝着赛博朋克的方向狂奔了。
Published:
我用了二十多年才培育起来的主体性，这些十七八岁的留子，就因为在 Stanford，在美国，或许更早在上海读了高中，就轻易许多地拿到了。
Published:
从library出来，有位头发花白的白人老奶奶看到我们，说了三遍excuse me。她非常proudly和我们讲话，掏出手机，说，看到你们俩在和这里的青蛙雕塑拍照。每逢圣诞节，这个青蛙都会穿上圣诞套装噢~瞧，就是这样！然后很真诚地问我们会不会在这里待到圣诞，问我们怎么知道Atherton这个地方的，因为这里实在是太不起眼、太静谧、太低调了。我俩说，可能要各回各家了。老奶奶又说她儿子从小不点的时候开始就学中文了，说的可六。
Published:
其实在从 SFO 回家这一路上，我挺伤心的。
Published:
今天在 Palo Alto 的 University Avenue 走了一圈，我反而想通了一件事，我们很多中国留学生，把硅谷当成了一个目的地。
Published:
Published:
Published:
发紧急邮件求教的时候，用 quick question 比写 emergency 更容易收到回复。
Published:
前几天杨立昆（Yann LeCun）在 LinkedIn 上发帖，说自己担任了一个新的职位，这在国内是几乎不可以想象的。
Published:
我意识到，我之前对于上课无用论，有一个非常严重的误解。
Published:
工作中突然冒出一件紧急的事，又没有时间准备，只能熬夜，这其实就是一种黑天鹅。
Published:
想要直观的感受到自己成长，就每天写日记。
Published:
我觉得高中以后的上课 a.e.（almost everywhere，几乎处处）没有任何意义。
Published:
闲聊历史，美国理科生比商科、法学、医学要给的 stipend 多，因为理科生穷。
Published:
现在越来越觉得，只有做生意，只有商业，只有贸易，才是最根本的慈善。
Published:
不要想第一条视频就能爆款，不要想第一次出去找生意就一定能拿到一千万，但是你可以想想，100 次之后会怎样。
Published:
为什么有些人看起来没你努力，却比你成功得早？因为成功首先是一个时间问题，然后才是能力问题。
Published:
今天又是高强度网上冲浪的一天，学到最多的一点是，世界上没有躺着就能财富自由的投资策略。
Published:
为什么不直接搞 2000 万，3000 万？
Published:
一杯奶茶，小费明明选的是 0，店家照样给你强制加 18%。
Published:
我刚刚在想，为什么我对证券的客服还挺客气，对中介却不能客气。
Published:
在学校里，考试就那么几次，你根本没有多少机会练。
Published:
我越来越鄙弃和蔑视「学习」这个词。
Published:
今天回看之前的日记，从物理的核聚变，到人类的把戏，无论一个 startup 还是一个行业，有一点很重要，就是自发涌动乃至永动的力量。
Published:
2024 年，在 40 余位学者中，我是唯一受邀的本科生。
Published:
The most terrifying thing about Silicon Valley has never been the brilliant ideas of geniuses, but that it established an entire system of mechanisms that encourage “creative betrayal.”
Published:
硅谷最可怕的从来不是那些天才的想法，而是它建立了一整套鼓励「创造性背叛」的机制。
Published:
搞科研就是赌博。事实上，没钱的人搞科研是赌博，搞什么也是赌博。
Published:
The most terrifying thing about Silicon Valley has never been the brilliant ideas of geniuses, but that it established an entire system of mechanisms that encourage “creative betrayal.”
Published:
硅谷最可怕的从来不是那些天才的想法，而是它建立了一整套鼓励「创造性背叛」的机制。
Published:
搞科研就是赌博。事实上，没钱的人搞科研是赌博，搞什么也是赌博。
Published:
摆摊的人赚的是差价，做成企业的人赚的是估值。两者之间隔着一个问题：谁能先拿到别人的钱。
Published:
由神经形态计算和基因语言模型驱动的虚拟细胞模型，代表了生物技术 AI 领域中资本最密集、科学最雄心勃勃的融合方向之一。如果你正在以 VC 实习生的身份评估这个领域，你需要理解三个核心组成部分：这些系统实际上做什么、为什么市场现在开始行动，以及可投资的机会在哪里。
Published:
建立在 19 世纪学徒制模型上的博士培养体系，几乎在所有关键经验指标上都已经出现了系统性失灵。
Published:
今天回看之前的日记，从物理的核聚变，到人类的把戏，无论一个 startup 还是一个行业，有一点很重要，就是自发涌动乃至永动的力量。
Published:
Although “umami” is the unshakable soul of Jiangsu-Zhejiang and Cantonese cooking, in traditional American food you often wander only among single-note salt, sweet, and oil, struggling to find that layered depth of flavor. That isn’t accidental; it’s the inevitable result of a deep cultural difference in how ingredients are handled, how seasoning is reasoned, and how industrial production works.
Published:
虽然“鲜”（Umami）在江浙菜和广府菜中是不可撼动的灵魂，但在传统美国菜里，你往往只能在单一的咸、甜、油之间徘徊，而难觅那种富有层次感的味觉深度。这并非偶然，而是一场关于食材处理、调味逻辑与工业化生产方式的深层文化差异所导致的必然结果。
Published:
Although “umami” is the unshakable soul of Jiangsu-Zhejiang and Cantonese cooking, in traditional American food you often wander only among single-note salt, sweet, and oil, struggling to find that layered depth of flavor. That isn’t accidental; it’s the inevitable result of a deep cultural difference in how ingredients are handled, how seasoning is reasoned, and how industrial production works.
Published:
虽然“鲜”（Umami）在江浙菜和广府菜中是不可撼动的灵魂，但在传统美国菜里，你往往只能在单一的咸、甜、油之间徘徊，而难觅那种富有层次感的味觉深度。这并非偶然，而是一场关于食材处理、调味逻辑与工业化生产方式的深层文化差异所导致的必然结果。
Published:
Published:
金涨多了就换成银，银涨多了再换回金，这个听起来像民间偏方的操作，其实是贵金属市场里最经典的增厚策略之一。
Published:
工作中突然冒出一件紧急的事，又没有时间准备，只能熬夜，这其实就是一种黑天鹅。