François Chollet: ARC-AGI-3, Beyond Deep Learning & A New Approach To ML

AGI expected by 2030

Chollet estimates AGI will likely emerge around 2030, coinciding with the release of ARC-AGI 6 or 7, and emphasizes that AI progress is unstoppable and accelerating.

Ride the acceleration wave

Rather than attempting to slow AI development, the critical question for builders is how to leverage and harness this unstoppable acceleration effectively.

Program synthesis at the foundation

NDIA is developing a new machine learning substrate based on program synthesis that operates below the level of coding agents, replacing parametric curves with minimal symbolic models.

Symbolic descent vs gradient descent

The lab uses 'symbolic descent' to find the simplest possible symbolic models explaining data, adhering to the minimum description length principle for better generalization.

The case for divergent research

Chollet argues that while the industry consensus focuses on scaling LLMs, pursuing alternative approaches like his—despite only a 10-15% chance of success—is essential because no one else will.

Coding agents exploit verifiable rewards

The recent breakthrough in coding AI succeeds because code provides formally verifiable reward signals (unit tests, compilation), allowing models to generate trusted training data autonomously.

The verifiability divide

Domains with formal verification like mathematics will be rapidly automated, while fuzzy domains like essay writing will see slow progress due to reliance on expensive human annotations.

Inefficiency of current stack

While LLMs could theoretically simulate AGI with sufficient compute, Chollet argues this would be profoundly inefficient compared to future optimal approaches operating at lower levels.

Defining true AGI

Chollet defines AGI not as economic automation but as human-level skill acquisition efficiency—the ability to master new tasks with minimal data like humans do.

Origins of the benchmark

He created ARC-AGI after discovering in 2016 that gradient descent could not learn generalizable reasoning algorithms, instead overfitting to surface patterns rather than discovering underlying programs.

Evolution to ARC-AGI-3

ARC-AGI V1 was too difficult for early models, V2 is now saturating, and V3 continues to measure true generalization as the field advances toward human-level sample efficiency.