Y Combinator · Video summary

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

François Chollet discusses a new AI paradigm beyond deep learning, focusing on symbolic models and program synthesis to achieve more optimal and efficient AI, with the ARC-AGI benchmark driving progress.

Y Combinator · March 27, 2026 · 57-min episode · 8-min read

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

Key takeaways

💡Current AI progress is inevitable and accelerating; the key question is how to leverage and ride the wave of AI advancements.00:09
🎯NDIA is developing a new branch of machine learning based on symbolic models and symbolic descent, aiming for greater optimality and efficiency than deep learning's parametric approach.01:08
🚀Explore the NDIA website and research papers to understand the principles of symbolic models and program synthesis as alternatives to deep learning.01:08
🔑The Minimum Description Length (MDL) principle suggests the shortest model of data is most likely to generalize; this is achievable with symbolic models, unlike parametric deep learning.04:11
🎯Consider how verifiable reward signals apply to your domain; focus on problems where solutions can be formally verified to leverage current AI advancements.06:34
🚀Progress in AI domains is limited by the availability of verifiable reward signals; domains like code and mathematics, which have natural verification mechanisms, are advancing rapidly.06:34

Chapters

1:08🚀 Introduction to NDIA and a New AI Paradigm
1:55💡 Program Synthesis vs. Code Generation
3:04📈 Symbolic Models vs. Parametric Curves
4:39⚖️ The Case for Alternate Approaches
6:34✅ Verifiable Rewards and Automation
9:52🧠 Defining General Intelligence (AGI)
11:41⏳ LLMs and Sample Efficiency
12:55🛠️ Origins of Keras and Early Deep Learning Beliefs
14:13📉 Limitations of Gradient Descent for Reasoning
15:35🖼️ The Genesis of ARC AGI
17:55📊 ARC V1 and the Advent of Reasoning Models
19:53🔄 ARC V2 and Post-Training Paradigms
26:08🎮 Introducing ARC AGI V3: Agentic Intelligence
29:38🕹️ Designing ARC V3 Games
34:19🧠✨ Fluid Intelligence vs. Knowledge
35:35💾 NDIA's Vision for Optimal AI Size
38:50🔬 Science as Symbolic Compression
40:13🌱 Learning Principles vs. Biological Plausibility
42:11🏗️ Building Compounding Foundations
44:11🎯 The Moving Target of ARC Benchmarks
46:13🔮 Timeline for AGI
46:38💡✨ Exploring Alternative AI Approaches
48:38📚 Learning from Early AI Research
50:22📈🚀 The Importance of Scalable Approaches
51:41🌐 Leveraging Existing Human Effort
52:57📜 Lessons from Growing Keras
55:44🌟 Advice for Aspiring AI Professionals

The digest

A New Paradigm for AI: Beyond Deep Learning and Towards Optimal Intelligence

This discussion explores a fundamental shift in Artificial Intelligence research, moving beyond current deep learning paradigms towards a more optimal and efficient approach. François Chollet, founder of the ARC Prize and NDIA lab, argues that while current AI progress is impressive, it relies on an inherently inefficient foundation. He proposes a new direction focused on symbolic program synthesis and a deeper understanding of intelligence as skill acquisition efficiency, aiming to build AI systems that are more data-efficient, interpretable, and capable of true generalization. This conversation is crucial for anyone involved in AI development, research, or strategic planning, offering a glimpse into the future of AI and a framework for thinking about its limitations and potential.