François Chollet: Why Scaling Alone Isn’t Enough for AGI

Chollet describes Indium’s work on program synthesis as an alternative foundation for machine learning, rather than a new kind of coding agent. Its approach replaces learned parametric curves with concise symbolic models and uses “symbolic descent” to search for them. He argues that shorter models could need less data, run more efficiently, and generalize better.

Chollet says coding agents improved quickly because code can be tested and rewarded with reliable signals, such as unit tests. The same dynamic could speed progress in mathematics, where proofs can be checked, while essay writing may advance more slowly because it depends on costly human judgments. He distinguishes task automation from AGI, which he defines as learning new tasks with human-like efficiency.

Chollet traces ARC’s origins to his effort to measure reasoning and skill acquisition. He says scaling base language models by 50,000 times did little to improve ARC-1, while reasoning models produced a sharp jump. ARC-2 then saturated through large-scale cycles of generating tasks, solving and verifying them, and training on successful reasoning chains.

ARC-3 shifts from solving static patterns to exploring unfamiliar interactive environments. Agents must discover controls and goals, build a model of the environment, and solve games efficiently; Chollet says human testers can do so without prior training. The benchmark’s private games aim to limit gains from training directly on the public set, and performance is scored for action efficiency.

At Indium, Chollet says the team spent about half a year finding foundations for a program-search system that can build on its own progress. He argues that AI approaches should improve without human engineers becoming the bottleneck, and predicts AGI around 2030 or the early 2030s. For open-source projects, he credits Keras’s simple API, helpful documentation, onboarding, and community building.