Y Combinator Startup Podcast François Chollet: The ARC Prize & How We Get to AGI
278 snips
Jul 3, 2025 
François Chollet, creator of the Keras library and founder of the ARC Prize, dives deep into the challenges of achieving true general intelligence in AI. He critiques existing models for their limitations and discusses why simply scaling up isn’t enough. Chollet emphasizes the need for adaptability and fluid intelligence over static skills. He introduces the ARC Prize as a benchmark for progress and shares insights from his research at Ndea on advancing AI's capabilities through abstraction and compositionality, laying a path towards AI that can reason and innovate.
AI Snips
Chapters
Transcript
Episode notes
Scaling Alone Fails ARC
- ARC’s failure to improve with scaling disproves fluid intelligence emergence by scale alone.
- Test-time adaptation is essential for genuine fluid intelligence in AI.
ARC 2: A More Sensitive Measure
- ARC 2 advances measurement by probing compositional generalization that static reasoning lacks.
- Current AI models show limited progress and remain far from human-level in ARC 2.
ARC 3 Focuses on Agency
- ARC 3 will test AI agency with interactive, goal-setting tasks in unknown environments.
- Efficiency in solving tasks will be required, simulating real-world learning challenges.