Jiayi Weng’s Podcast Went Viral, So I Watched All of It

Jiayi Weng planned his career at OpenAI around one goal, to get his name onto as many OpenAI blog posts as possible.

Author: Koutian Wu; GitHub: ktwu01

His episode on the WhyNotTV podcast went viral, so as someone who works in AI I watched the whole thing and took notes.

As a kid he disliked Chinese class and loved math olympiad, solving problems faster than anyone, and later made it to the national olympiad in informatics. At Tsinghua he attacked an information gap by open sourcing his homework. A line still goes around, “Do you know so and so? No. Do you know Jiayi Weng? Yes, I survived on his homework.” He also hacked the site the school used for downloading transcripts and cut the cost to one hundredth.

He found reinforcement learning, which he defines in one line as machine learning with feedback from an environment, through a ViZDoom project. That led to a summer research position and to building RL infrastructure in his senior year. His summer research was on mixture of experts. He spent a long time learning transformers and NLP, but he lacked the engineering and the compute, and no first author paper came out. When he applied for PhDs, even with a letter from a Turing Award winner, the season left him a bit defeated.

Back then the pecking order at Tsinghua made him think a PhD beat a master’s. But he wanted out of that scoring system, to do something different that he actually liked instead of chasing GPA. As an undergraduate he used the yardstick of his advisor Hang Su, papers, competitions and GitHub stars, so he started contributing to open source, which compounds over time. In his master’s he learned that employers care about relevant experience, not GPA. What matters most is whether you can deliver.

During COVID he insisted on going abroad and kept writing two projects. One was Tianshou, an RL library on GitHub that later passed 9k stars. The other was tuixue.online, which tracked open US visa appointment slots. Why Tianshou? He had already written a lot of infra code, and the existing RLlib was too complex to use, so he built his own, alone, with a first version in two weeks. He thinks code consistency matters a lot and that more people can rot it. Both projects came from his own needs, and so did collecting old homework. He open sourced them as charity, not for money but for impact. Since high school he had pictured the end screen of his life and wanted to create value for others. Impact is not fame.

He got a CMU offer, plus offers from the company that became DeepSeek, OpenAI, NVIDIA and TikTok, and chose OpenAI to see how the most advanced RL research in the world gets done. He made that choice with no idea ChatGPT was coming.

He joined John Schulman’s team. Why did John hire him? Mostly because his GitHub looked great, and John respected his yardstick and his engineering. The interview was an end to end problem with a three hour limit. He finished in two.

His view is that doing a PhD and then going to industry wastes your life. If you build up citations and projects during undergrad or a master’s, reaching PhD level papers or open source infra with real engineering and something distinctive, you do not need the PhD. He also dislikes chasing funding. He quoted a colleague who said teaching researchers to do engineering is far harder than the reverse, because research ideas are cheap and engineering, infra and iteration matter more. Every LLM has bugs, and whoever fixes them fastest moves fastest.

The part that struck me most was his career plan. Be the founder of the foundation. Build infra so others star it, train on it, cite it and add your name. He built OpenAI’s entire RL infra for post-training, and his name is on every release. RL sits near the top of the infra stack, so everyone who uses it considers crediting him, and he planned it that way, because a single research project cannot scale. His advice is that academia will be restructured in the long run, and if you want to get into infra, start early. Infra is a bottomless pit and very few people have stuck with it for more than three years, which is why in 2022 he went all in on RL infra instead of Tianshou.

When he joined OpenAI, GPT was not the main focus and there was no split between pre-training and post-training yet. Because he watched it improve step by step, he did not see how disruptive GPT was, while outsiders seeing it for the first time were stunned. When OpenAI’s servers got crushed by traffic, it felt like tuixue.online getting crushed years earlier, the users pouring in like a tide. His first impression of OpenAI was a big lab, with many strong researchers pointing the way and very high talent density.

On trial and error he cited Google’s productivity lesson. The faster you try, the higher your success rate. You do not need many genius ideas, you need fast iteration. OpenAI grew from 280 people to 3,000 and stayed efficient, with information flowing up and down. Organizing a company and organizing a repo both need strong consistency.

He talked about the cost too. He overworked, six days a week from morning to night, until a headache sent him to the emergency room. That is when he realized he had to protect his health to last, and now he runs three kilometers twice a week.

He thinks the biggest bottleneck for LLMs now is the throughput of fixing infra and debugging, the speed of iteration per unit of time. His team is rebuilding OpenAI’s infra from scratch to clear technical debt. He also thinks OpenAI researchers may be the first job AI replaces, because ideas are too cheap, while sales is harder to replace because it is a game between people. On AGI, the joke inside OpenAI is that twenty people have twenty definitions. Nobody dares to let AI change the infra yet, so by his own definition AGI is not here.

Asked whether OpenAI still has anything to do with open, he called it a tradeoff. You cannot open source your best model and survive. Benefiting all of humanity through AGI comes in two steps, first scale up and reach AGI, then deliver the benefit. The openness is meant for ordinary people later, not for competitors now. With unlimited resources, would he open source? Of course. DeepSeek releasing its weights made OpenAI reconsider.

At the end he said he used to love building RL infra but now feels a little lost, because he can see the end of the road and AGI looks certain. In ten years he hopes to have enough resources to live the life he likes. His last words to the podcast were, go find out what you want most.