The Future World: AI Assistants, Collaboration, and the End of UI Friction

In the future world, everyone will have their own AI agent assistant.

Author: Koutian Wu; GitHub: ktwu01

When you look at physicists, their ultimate goal is to understand how to predict the world, how to understand the ultimate questions of existence, and how humanity lives. All scientists share this goal of understanding the world. But physicists take it a step further—they believe that high-dimensional worlds have low-dimensional manifolds, and they use classic learning to manifest them. So, how can an ordinary Ph.D. student, with just an hour or two of free time a day, participate in these grand questions? The most direct way might be to join places like Google DeepMind.

Does this mean we won’t need to learn physics anymore, and AI will do all the derivations? Even if AI does the deriving, humans will still need to learn those concepts. The things we have to learn will only increase and get harder. The nature of the difficulty will change. Before, the difficulty was doing arithmetic by hand; now, with computers, we don’t need slide rules, but the difficulty has shifted to a different dimension.

What becomes crucial in this future is large-scale, crowdsourced collaboration. We recently saw this when a group of us crowdsourced a research paper, with each person taking a section. Soon, the scope of collaboration will expand beyond writing articles to creating movies or even massive 3D games like GTA. AI is empowering this era, making these massive creative endeavors accessible.

But to truly unlock this future, we have to rethink how we interact with AI. As I wrote in If Jobs Were Building Today, He Would Toss Most multiAgent Products in the Trash, the current state of multi-agent platforms is flawed. They show you every step—calling weather tools, routing to agents, narrating the entire process.

Steve Jobs understood that all of the complexity should be hidden inside. You hand the AI your dirty clothes, and it brings them back clean. You don’t need to know how it hailed a cab or which laundromat it used.

The UI itself has become the friction. In the future world, forced prompt engineering and bouncing between clunky SaaS frontends will be legacy clutter. AI agents won’t need beautiful dashboards; they will need APIs and structured endpoints. The product that wins won’t be a complex toolbox you have to configure; it will be an invisible layer that just magically works.

When the friction of execution disappears, humanity’s bandwidth is freed up to tackle the real challenges: predicting the world, understanding the universe, and building unprecedented creative collaborations together.

Everything involving humans becomes distributed

Zoom out one level: everything becomes distributed. Companies, learning, collaboration, anything that involves humans is being pulled apart from a closed, centralized structure into a loose, distributed network.

A few concrete cases:

1. Writing an article or paper. It is no longer one person staring at a blank document until something happens. It is a group of people plus a group of agents collaboratively assembling the result. Each section, each literature review pass, each figure can be handed off to a different human or agent and merged back in.
2. Everything else around it. Review, data wrangling, experiment execution, reproduction, figure generation, and code: the same pattern applies. The workflow gets sliced into smaller nodes, the nodes get distributed, the results get merged back together.
3. Large-scale collaboration itself is changing shape:
   • (a) A PhD is becoming a distributed task. It is no longer “one student + one advisor + one room.” It is a cross-institution, cross-geography network task with multiple humans, multiple agents, and multiple data sources all contributing.
   • (b) A paper can have many authors; one person’s PhD can be supported by many people. The advisor, collaborators, agent assistants, remote mentors, and the broader community all contribute to the making of the same PhD.

You can see the same pattern outside academia. A film is less and less just one studio working behind closed doors for three years; it is distributed collaboration plus an AI toolchain plus a global remote team. In academic publishing, the visible signal is undergraduates publishing research, because the minimum unit of “publishable work” has been sliced small enough that an undergrad plus the right agents plus the right collaboration network can assemble a publishable paper.

There may not be more arbitrageurs, but the good ones get much stronger

Distribution does not automatically turn everyone into a winner. The number of arbitrageurs will not necessarily go up, because human nature does not change. Most people will not suddenly start tackling harder problems just because the tools got cheaper. Most people will keep doing the same comfortable thing.

But the people who actually exploit the arbitrage will become dramatically more effective. The tools they can call, the collaboration networks they can stitch together, the number of tasks they can push forward in parallel, all of these are amplified by an order of magnitude. The result is winner-takes-all: a small number of people capture most of the upside, and everyone else falls further behind the curve.

Startups are the same story from a different angle. The bar to start a company keeps dropping. One person plus a handful of agents can stand up an MVP and validate an idea at near-zero cost. So more and more people are starting companies. But a lower bar does not mean a higher survival rate; most of them still die, they just die faster and cheaper. The few that win will win more completely than the last generation of founders.

A real example: a project involving a high-schooler submitted to Science, a CS professor saying he has no edge over non-coders

This is not abstract speculation, it is already happening. Chuanfei Dong, BU professor and Scientific Committee member at earth-space-ai, told me recently:

Yesterday I was involved in a high-school student’s project that we submitted to Science [Lol], also AI-related. If it gets accepted, it would suggest that with AI, even a high-schooler can publish in Science now. The bar to do research has dropped a lot. I’m at this conference today, scifmconferences.org, and a CS professor at the conference said he no longer has any advantage compared to someone who doesn’t know how to code [Lol].

Those two lines point to the same shift from two angles:

• The bar has been lowered. Work that used to require years of training in research, programming, and publishing is now within reach of a high-schooler with the right tools.
• The gap at the top gets wider. A CS professor saying he has no advantage over a non-coder does not mean the professor got worse. It means the differentiator is no longer “can you write code.” It is “can you formulate a problem worth handing to an agent” and “can you orchestrate a distributed collaboration network.” Those are more upstream skills, and the spread between people on those skills is much wider.

Someone has to build the plumbing

Distribution is the direction, but it does not arrive for free. For humans + agents + cross-institution networks to actually collaborate reliably, someone has to write the A2A (agent-to-agent) protocol. Someone has to define how agents discover each other, authorize each other, exchange intermediate results, and verify provenance. Someone has to build cross-agent state sync, task hand-off, and accountability tracing. Someone has to solve the credit-assignment problem when a paper has 30 authors and many agents and you need to know who and what did which part.

That layer is mostly empty right now. Whoever builds it has a shot at occupying a position in the distributed era analogous to TCP/IP in the internet era. It is one of the highest-leverage things to work on today, and one of the biggest pieces of arbitrage still sitting there.