Thesis-Driven Entrepreneurship

The unit of a startup used to be the product. In the AI era, it is becoming the thesis.

This sounds abstract, but it is a practical change. A product is a thing you can launch. A thesis is a way of seeing the world that keeps producing things to launch.

For the last twenty years, the default advice to founders was product-driven. Find a painful problem. Build a product. Talk to users. Iterate. This was good advice, and it still is. But it assumed that the product was the main container of learning. You learned by improving the thing you had built.

AI makes that container less stable.

The model changes. The interface changes. The cost curve changes. What was impossible in March becomes obvious in October. A workflow that looked like a product one year becomes a feature of a foundation model the next. In ordinary software, a good product could be a durable object. In AI, a good product is often a temporary expression of a deeper view.

So the question for founders changes from “What product should I build?” to “What thesis can I live inside for ten years?”

By a thesis I do not mean a slogan. Most company missions are too vague to be useful. “Make AI useful” is not a thesis. “Democratize knowledge” is not a thesis by itself. A real thesis has edges. It says something about where impact is trapped, why other people have not unlocked it yet, and why a small team has an unfair chance to unlock it now.

A good thesis has at least three properties.

First, it points at unrealized impact. There is some large class of work the world wants to do, but cannot yet do well. Not because people are lazy, but because the tools, economics, or coordination mechanisms are wrong.

Second, it contains asymmetric information. The founder has seen something that the market has not priced in. This may come from research taste, domain experience, distribution insight, or simply from living closer to the future than other people.

Third, it can generate many products. If the first product fails, the company should become wiser, not empty. The failed product should teach the thesis where reality is harder or more interesting than expected.

This is what makes thesis-driven entrepreneurship different from merely having a vision. A vision can be decorative. A thesis is operational. It tells you what to try next.

You can see this pattern in the new AI companies that feel most like the future.

MiroMind does not describe itself mainly as a chat app. It says it is building a “General Purpose Solver” and “Discoverable Intelligence,” with systems that run a loop of Plan -> Execute -> Verify -> Improve until an answer is not merely plausible but validated. Its open-source MiroFlow and MiroThinker work are products, but they are also probes into a larger thesis: that the next important AI systems will be reasoning systems with verification built into the loop.

Whether MiroMind succeeds is a separate question. The interesting thing is the shape of the company. If the mobile app fails, the thesis is not dead. If a benchmark becomes obsolete, the thesis is not dead. The company can try research agents, scientific reasoning tools, financial intelligence systems, or verifiable software agents. The product surface can move because the thesis is deeper than the surface.

Thinking Machines Lab is similar in a different direction. Its public statement says AI systems should become more widely understood, customizable, and generally capable, and that research and product should be co-designed. Its first product, Tinker, is not “a better chatbot.” It is a training API that gives researchers control over model training while the company handles the infrastructure.

That is a thesis in product form. The thesis is that the frontier is not only bigger models, but more people being able to adapt models to their own goals. Tinker is one implementation of that view. It may turn out not to be the final one. But if you believe the bottleneck is customization, not just raw intelligence, then many products follow: training APIs, evaluation tools, data environments, collaborative model labs, domain-specific adaptation systems.

Anthropic is an older and more obvious example. The company says it wants to build reliable, interpretable, and steerable AI systems. Claude is a product, but it is also an embodiment of that thesis. The Model Context Protocol is another. It says, in effect, that useful AI will need a standard way to connect to the systems where work actually happens.

The important point is not that every thesis has to be about safety or protocols. It is that Anthropic can ship many different products without becoming incoherent, because they point back to the same belief about what AI systems must become.

Cursor may be the cleanest product example. It began as an AI code editor, but the company increasingly talks about a third era of software development: developers directing fleets of agents, reviewing artifacts, and building the factory that builds software. Cursor 3, cloud agents, rules, skills, and the agent harness are not random features. They are successive attempts to express the same thesis: the programmer’s job is moving upward, from editing lines to orchestrating agents.

This is why “Cursor for X” is both useful and misleading. The interesting part of Cursor is not that it put AI into an existing product category. Many people did that. The interesting part is that it rebuilt the workflow around the agent. If you copy the surface, you get a feature. If you understand the thesis, you get a company.

Cognition’s Devin makes the same point from another angle. Cognition says it builds tools so engineers can operate more like architects while agents handle repetitive engineering work. Devin is the product. The thesis is broader: software creation is becoming delegable, auditable, and parallelizable. From that thesis you can get coding agents, migration agents, on-call agents, agent teams, enterprise automation, or tools for supervising all of them.

Again, the product matters. A startup cannot live on a thesis alone. Users do not pay for a philosophical position. They pay for a thing that helps them. But in AI, the thing may have a short half-life. The thesis is what lets the company survive the half-life.

This also explains why some AI products feel impressive but not like startups. They are demos without a thesis. They show that a model can do something surprising, but not that the founder understands a durable wedge. The demo may get attention. It may even get revenue. But when the model provider ships the same capability, or when the novelty fades, there is nowhere to go.

The danger of product-driven thinking in AI is that it can make founders optimize for the first visible application. The first visible application is often too close to the model’s current weakness. A founder sees that agents are bad at a workflow, builds scaffolding around that weakness, and calls it a company. Six months later the weakness disappears. The product was real, but the thesis was just “models are bad at this today.”

That is not enough.

A better thesis is about what becomes possible as models get better. Cursor does not need models to remain weak at coding. It benefits when they become stronger, because the job to be done expands from completing code to coordinating software production. Thinking Machines does not need models to remain hard to fine-tune. It benefits if more people want to adapt them. MiroMind does not need reasoning to remain bad. It benefits if verified reasoning becomes more valuable as reasoning systems enter higher-stakes domains.

The best AI theses ride capability curves instead of hiding from them.

This is where Paul Graham’s old advice becomes more important, not less. He wrote that startup ideas are not blueprints but questions, and that founders should be willing to change the idea after users teach them what is true. In AI, the product idea is even more obviously a question. But the thesis is the frame that decides which questions are worth asking.

If you have no thesis, iteration can become random. Users ask for something, you build it. A competitor launches something, you react. The model improves, you pivot. Eventually the company becomes a pile of clever reactions.

If you have a thesis, iteration compounds. Each product is an experiment in the same direction. Each failure narrows the map. Each customer conversation teaches not just what users want, but whether your view of the future is becoming more or less true.

This is also why venture capital is naturally thesis-seeking in AI. A venture-scale company needs more than a useful tool. It needs a widening market. It needs a reason that early learning will become more valuable over time. The investor is not just asking, “Is this product good?” They are asking, “If this product works, what else becomes possible?”

The founder should be asking the same question.

There are bad versions of thesis-driven entrepreneurship. The most common is using the word “thesis” to avoid building. A founder writes a beautiful memo about the future and never finds a user. That is not thesis-driven. That is essay-driven.

The other bad version is refusing to update. Some founders become attached to their thesis the way bad founders become attached to their product. But a thesis is not a religion. It is a compression of what you have learned so far. It should be strong enough to guide you and weak enough to be improved by reality.

The practical test is simple.

Write down your current product. Then imagine it fails for a specific reason. The user need was smaller than expected. The distribution channel was too expensive. The model provider absorbed the feature. The enterprise buyer would not trust it. The workflow was too political. Whatever the reason is, make it real.

Now ask: after that failure, do you still know what to try next?

If the answer is no, you did not have a thesis. You had a product idea.

If the answer is yes, ask the next question: did the failure make the thesis sharper?

That is the difference. Product-driven entrepreneurship asks whether the product worked. Thesis-driven entrepreneurship asks whether the company learned something that can generate the next product.

This does not mean founders should become less concrete. The opposite is true. A thesis only matters if it keeps forcing concrete attempts. The product is the proof that the thesis is not just language. The user is the judge. Revenue is evidence. Retention is evidence. A painful deployment is evidence. A failed launch is evidence too, if it teaches you where the thesis was wrong.

The old startup question was: can you make something people want?

The new question is: can you keep making things people want as the frontier moves?

In stable markets, a great product could be the answer. In AI, the answer is more likely to be a great thesis, expressed through a sequence of products.

The product is an experiment. The thesis is the laboratory.