Scaling the current generation of deep learning models is an inefficient and insufficient path to AGI, as it primarily enhances pattern matching rather than true reasoning or fluid intelligence.
True general intelligence must be measured by a system's ability to efficiently learn and become competent at any novel task, a principle embodied in the ARC-AGI benchmark.
AGI is fundamentally an algorithmic problem that likely has an elegant, compact solution (e.g., <10,000 lines of code), rather than one requiring incomprehensible scale.
Program synthesis offers a more promising foundation for future AI systems than parametric deep learning, promising greater data efficiency and generalizability.
Current LLM-based technology can fully automate any problem domain that has formally verifiable solutions and a trusted reward signal, such as mathematics and coding.
2016
While at Google Brain, Chollet's research indicated that gradient descent was insufficient for discovering generalizable algorithms for reasoning, instead defaulting to pattern matching.
2019
Chollet published his paper "On the Measure of Intelligence" and created the first version of the ARC-AGI benchmark, containing 800 tasks he personally designed to test fluid intelligence.
Circa 2020-2023
Observed that base LLMs, including early models like GPT-3, scored zero or near-zero on ARC-AGI V1, reinforcing his view that scaling pre-training alone was not improving fluid intelligence.
Present
Founded the research lab Endia to develop a new AI substrate based on program synthesis. Simultaneously, he is advancing the ARC-AGI benchmark to V3 to measure "agentic intelligence" in interactive, game-like environments.
Future
Plans to release future versions of the ARC-AGI benchmark, with V4 focusing on continual learning and V5 on invention, while predicting AGI will be achieved in the early 2030s.
▶Critique of the Scaling ParadigmApr–May 2026
Chollet consistently argues that increasing the scale of deep learning models is not a viable path to AGI. He supports this with evidence from his 2016 research on gradient descent and the poor performance of base LLMs on the ARC-AGI benchmark, even after a 50,000x increase in model scale.
Investors heavily allocated to companies whose primary competitive advantage is scale of compute and data for training LLMs face a significant thesis risk if Chollet's view that a different, more efficient paradigm is required for AGI proves correct.
▶Measuring and Defining 'True' IntelligenceApr–Jul 2026
A core part of Chollet's work is defining and measuring a more general, human-like intelligence. He operationalizes this through the ARC-AGI benchmark series, which evolves from static reasoning puzzles (V1) to interactive, game-like environments that test "agentic intelligence" and skill acquisition (V3).
The ARC-AGI benchmark acts as a leading indicator for the future direction of AI research; its increasing adoption could shift focus and funding from pure scaling towards models that demonstrate efficient learning and adaptation in novel environments.
▶Pioneering Program Synthesis as an AlternativeApr 2026
Through his lab, Endia, Chollet is developing a new machine learning substrate based on program synthesis. This approach, which includes methods like "symbolic descent," aims to create models that are more data-efficient and generalizable than current parametric deep learning systems.
Endia represents a high-risk, high-reward venture that could fundamentally disrupt the AI hardware and software stack, as its success would favor different computational models than the matrix multiplication-heavy architecture that currently dominates the industry.
▶AGI as an Elegant, Algorithmic ProblemApr 2026
Chollet frames the pursuit of AGI not as a problem of brute-force scale, but as a search for the right algorithm. He speculates that the final AGI solution will be a surprisingly small codebase (under 10,000 lines) and that the core concepts could have been implemented with 1980s-era computing resources.
This perspective suggests that the winner in the AGI race may not be the entity with the most GPUs, but rather a small, focused team that discovers the correct algorithmic breakthrough, potentially leveling the playing field for smaller research labs and startups.