Y Combinator Startup Podcast How François Chollet Is Building A New Path To AGI
120 snips
Mar 27, 2026 
François Chollet, AI researcher behind Keras and founder of ARC Prize and Ndea, explores a radically different route to AGI. He gets into symbolic programs over neural nets. He talks about why coding agents are suddenly thriving, where deep learning may hit limits, and how ARC V3 tests learning, planning, and adaptation in new environments.
AI Snips
Chapters
Transcript
Episode notes
ARC Predicted Reasoning Models and Coding Agents
- ARC V1 exposed that scaling base LLMs was not enough, while reasoning models produced a sudden jump in performance.
- ARC V2 then tracked the rise of agentic coding, where targeted post-training rapidly saturated the benchmark.
Harnesses Increase Usefulness More Than Intelligence
- François Chollet says ARC V2 was saturated by RL loops and custom harnesses, not by a big jump in raw intelligence.
- Models got more useful through synthetic tasks, verified solutions, and execution-aware training, while humans still supplied the solving strategy.
ARC V3 Tests Exploration In Unknown Worlds
- ARC V3 measures agentic intelligence by dropping an AI into unfamiliar mini-games with no instructions, goals, or control map.
- Success depends on human-like exploration efficiency, world modeling, goal inference, planning, and execution within a limited action budget.