Y Combinator• Mar 27, 2026• 57:23Interview

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

From Y Combinator

François Chollet•Founder, Endia; Creator of Keras

Executive Summary

François Chollet predicts AGI will likely be achieved in the early 2030s, but argues the current deep learning stack is inefficient and will not be the long-term foundation for AI.
Chollet's new lab, Endia, is developing an alternative AI substrate based on program synthesis and "symbolic descent," aiming for more data-efficient and generalizable models.
The recent explosion in AI capabilities, especially in coding, is attributed to the presence of formally verifiable reward signals, a principle that will soon revolutionize other domains like mathematics.
The ARC-AGI benchmark has evolved to V3, which measures "agentic intelligence" in novel, game-like environments to better test for true fluid intelligence and prevent "teaching to the test."

12 quotes

Concerns Raised

The current deep learning/LLM stack is inefficient and not the optimal long-term path to AGI.
Scaling existing models alone is insufficient for achieving general intelligence, as demonstrated by early ARC benchmark results.
AI progress in many domains is limited by the lack of formally verifiable reward signals, unlike in coding or mathematics.

Opportunities Identified

Developing alternative AI substrates like program synthesis could lead to breakthroughs in efficiency and generalization.
Any problem domain with a formally verifiable reward signal can be fully automated with current technology.
Individuals can leverage AI as a powerful tool for empowerment by combining it with deep domain expertise.

Key Themes

The Inefficient Path to AGI

François Chollet is bullish on achieving AGI by the early 2030s but is critical of the current deep learning and LLM paradigm. He argues that while this approach may succeed, it is fundamentally inefficient and suboptimal, and will likely be replaced by more elegant solutions in the long run.

This highlights a critical debate in AI research: whether the current path of scaling up existing architectures is sustainable or if a paradigm shift is necessary for more efficient and robust intelligence.

Program Synthesis as a New AI Substrate

Chollet's lab, Endia, is pioneering an alternative to deep learning based on program synthesis. This approach, using a method called "symbolic descent," seeks to find the simplest possible symbolic model to explain data, which should require less training data, be more efficient, and generalize better.

This represents a contrarian but potentially revolutionary research direction that could leapfrog the limitations of current parametric models, offering a path to more optimal AI.

The Power of Verifiable Reward Signals

The recent success of AI coding agents is not just due to better models, but because code provides a clear, verifiable reward signal (e.g., unit tests pass or fail). Chollet posits that any domain with such a signal, like mathematics, can be fully automated with current technology.

This provides a clear framework for identifying which industries and tasks are most susceptible to immediate automation, guiding business strategy and investment in AI applications.

Evolving Benchmarks to Measure True Intelligence

The ARC-AGI benchmark is continuously evolving to create a more accurate measure of general intelligence. While V1 and V2 were eventually saturated, the new V3 benchmark introduces interactive, game-like environments to test "agentic intelligence," focusing on exploration, goal-setting, and planning in unfamiliar contexts.

This evolution in benchmarking is crucial for the field to avoid optimizing for narrow tasks and instead drive research towards creating AI with genuine, flexible problem-solving abilities.

Get started free

Topics

François Chollet AGI Artificial General Intelligence AGI Timelines Endia Program Synthesis Symbolic Descent Deep Learning LLMs AI Benchmarking ARC-AGI Benchmark Agentic Intelligence Reinforcement Learning Reward Signals Task Automation AI Coding Agents Keras

Processed Apr 15, 2026 yt-dlp + mlx-whisper + Gemini