Keras creator launches Ndea to rethink AGI

François Chollet is done grading other people's homework.

After ten years at Google, the creator of the widely-used Keras deep learning framework and the creator of the brutal ARC benchmark is leaving to build his own AI lab. Chollet has teamed up with Zapier co-founder and ARC Prize partner Mike Knoop to launch Ndea, a research lab taking an unconventional bet on artificial general intelligence.

Instead of just scaling up transformers on massive text datasets, Ndea is focusing on deep learning-guided program synthesis—teaching AI to write executable code on the fly to solve novel problems.

Merging intuition with cold logic

Most modern frontier labs view program synthesis as a minor feature in a broader generative model. Ndea views it as the other half of the brain. In fact, the company believes program synthesis is sitting roughly where deep learning was back in 2012, just before its massive boom.

Even the lab's name reflects this philosophy. Ndea merges two Greek concepts: ennoia (intuitive understanding) and dianoia (logical reasoning). The goal is to marry the fuzzy pattern recognition of deep learning with the rigorous, explainable logic of program synthesis.

Chollet put it candidly on X: the team believes they have a small but real chance of achieving a breakthrough to create AI that learns at least as efficiently as humans do, without running into the scaling bottlenecks currently facing LLMs.

Turning side projects into the main event

For years, Chollet's foundational paper "On the Measure of Intelligence" and the ARC benchmark were essentially side projects at Google. With Ndea, they are taking center stage.

The launch comes at a fascinating moment. OpenAI recently flexed its upcoming o3 reasoning model on ARC, hitting 75.7% with standard compute and 87.5% with high compute. Chollet acknowledged the progress, but pointed out that o3 still stumbles on simple tasks and relies heavily on brute-force search rather than genuine human-like generalization.

Now, Chollet is transitioning from the benchmark creator to the solution builder. Backed by Knoop, who has already invested in AI firms like Cognition and Braintrust, Ndea is building a globally distributed, remote research team aimed at operating like a factory for rapid scientific advancement.

Why it matters

While the rest of Silicon Valley is pouring billions into training bigger models on more of the same internet text, Ndea represents one of the highest-profile bets on an alternative architecture.

For developers and researchers, there is nothing to test or sign up for today. But this shift signals where cutting-edge reasoning research is heading. The top-performing submissions on the ARC leaderboard recently have relied precisely on combining deep learning priors with program search. If Ndea succeeds, future developer agents might move beyond statistical autocomplete into real algorithmic problem-solving.

We think Chollet's move is both timely and gutsy. LLM scaling is increasingly facing data limits and compute costs, and program synthesis offers a principled way to ground model outputs in formal logic. That said, program synthesis has historically suffered from massive combinatorial search spaces. Marrying it with neural networks is notoriously hard to scale, and Ndea has a mountain to climb if it wants to outpace the brute-force progress of mainstream AI giants.