Francois Chollet: The Founder Playbook

Francois Chollet is the French engineer who created Keras, the open-source deep learning library that put neural networks in reach of ordinary developers, and later the ARC-AGI benchmark that reframed how the field measures machine intelligence. Founders study francois chollet because he keeps winning from the contrarian side of the table: he built the most-used API in deep learning by obsessing over usability, defined artificial general intelligence on his own terms, and left a Senior Staff role at Google after nine years to start Ndea, an AGI lab that bets against the industry's scaling consensus. He is one of the most useful tech founders to model precisely because his edge is not raw compute or capital, it is judgment about which unpopular bet is worth making. You can follow his thinking on his profile.

This playbook pulls his operating principles from his own talks and shows how to apply them if you are a founder building with AI today.

Take the asymmetric bet

Chollet's whole career runs on one rule: chase the idea that almost certainly fails but pays off enormously if it works, especially when nobody else is on it. In François Chollet: Why Scaling Alone Isn’t Enough for AGI, Francois says, "If you have a big idea and it is very low chance of success but if it works it's going to be big and no one else is going to be working on it, then you should try a chance."

That logic explains all three of his big moves. Keras was a side project when the field assumed serious researchers wrote raw TensorFlow. ARC-AGI was a benchmark for a kind of reasoning the leading models were nowhere near solving. Ndea is a bet that program synthesis, not more scaling, is the path to AGI, which is a deeply unfashionable position to hold in 2026.

So what: your defensible ideas are the ones your competitors think are a waste of time. If ten other startups are building the same obvious AI wrapper, the expected value is already priced out. Look for the bet where the downside is a few months of your time and the upside is a category nobody else is contesting.

Win on usability, not raw technology

Keras did not beat its rivals on benchmark numbers. It won because it was pleasant to use. Chollet attributes its success largely to a focus on usability and an intuitive API, the developer's actual experience of the tool rather than its underlying horsepower. When a dozen frameworks could train the same network, the one that got out of the developer's way became the default.

This is the most transferable lesson in his playbook for AI founders, because the raw model is increasingly a commodity everyone can rent. The moat moves to the surface the user touches: how fast they reach a result, how few concepts they have to learn, how rarely they hit a wall. The same shift shows up in how teams build software now, where the leverage is in the workflow around the model, not the model weights (see from vibe coding to agentic engineering).

So what: if your product and a competitor's both call the same foundation model, usability is your differentiation, so treat time-to-first-value and API ergonomics as first-class features, not polish you add later.

Turn your users into maintainers

Chollet did not scale Keras by hiring a large closed team. His advice for building a successful open-source project is to focus on community building and to hire your most enthusiastic users to help maintain it. The people who already loved the tool became the people who extended it, which compounds: every engaged user is a potential contributor, evangelist, or future teammate.

For a founder with AI, this reframes distribution as a byproduct of genuine engagement rather than a separate paid-acquisition line. A user who ships something real with your product and tells others is worth more than a click you bought, and your best early hires are often already in your community solving their own problems with your tool.

So what: instrument who is getting the most value from your product, then give them ways to contribute, whether that is a public roadmap, a contributor path, or a direct recruiting conversation. Your power users are your cheapest and highest-trust growth channel.

Ride the wave instead of fighting it

Chollet treats accelerating AI progress as inevitable, and he thinks the founder's job is to position for it rather than fear it. In François Chollet: Why Scaling Alone Isn’t Enough for AGI, Francois puts the framing bluntly: "How do you make use of it? How do you leverage? How do you ride the wave? That's the question to ask." He sees AI as empowerment and leverage, not primarily a threat to jobs.

The practical version of this for a builder is to assume the models will get cheaper and more capable on a schedule outside your control, then design so that curve helps you instead of stranding you. That means not overfitting to a single model's current quirks and keeping the parts you own, the data, the evaluation harness, the user experience, portable across whatever ships next.

So what: build a rigorous evaluation loop now so you can adopt each better model the day it lands without guessing whether it actually improved your product (see LLM evals for founders). The founders who ride the wave are the ones who can measure it.

Own the measuring stick

Chollet's most strategic move was defining the game before playing it. He defines AGI as a system that can approach any new problem and become competent at it with the same efficiency as a human, then published the ARC-AGI benchmark in 2019 to measure exactly that, and then co-founded Ndea to win by that definition. He set the industry's yardstick before competing on it, which is a position of enormous leverage: everyone else now argues on his terms.

His conviction behind Ndea is that scaling alone plateaus and that the field trends toward data-efficient approaches. In the same talk, Francois says, "I think it's inevitable that the world of AI will trend over time towards optimality," and he pegs a timeline to it: "I think we're probably looking at AGI 2030, around the time that we're going to be releasing like maybe Arc 6 or Arc 7." He is betting on concise symbolic models that need less data and generalize better than today's parametric ones. This is the same category-defining instinct that founders like Demis Hassabis show when they frame a research agenda the rest of the field then follows.

So what: whoever defines how success is measured in your category shapes how customers and competitors think. Publish your own benchmark, scorecard, or standard for the problem you solve, and you stop competing inside someone else's frame and start setting it.

What to copy this week

• List your current bets and mark each by odds and payoff; kill the crowded low-upside ones and put real time into the one contrarian bet with asymmetric upside.
• Audit your product's time-to-first-value and cut one step of friction between signup and the user's first real result.
• Identify your three most engaged users this week and give at least one of them a concrete way to contribute or a recruiting conversation.
• Stand up or tighten an evaluation harness so you can swap in the next model on release day and prove whether it improved your product.
• Draft the scorecard or benchmark that defines success in your category, then publish it so the conversation happens on your terms.

AI Operating System for Startups

Sources

• François Chollet: Why Scaling Alone Isn’t Enough for AGI
• Ndea: About
• TechCrunch: AI researcher François Chollet founds a new AI lab focused on AGI
• TIME 100 Most Influential People in AI 2024: François Chollet
• François Chollet: personal page
• Francois Chollet on X