ARC-AGI Explained: Closing the Human–AI Gap — François Chollet

Frontier Data Summit 2026: François Chollet explains how AI benchmarks should evolve to measure the remaining gap between human and AI capabilities. François Chollet joins Fred Sala to discuss AI benchmarks, ARC-AGI, and how to measure the remaining gap between human and AI capabilities. They explore creativity, open-ended environments, scientific insight, transfer learning, and whether training on verifiable tasks can lead to broader problem-solving ability. IN THIS VIDEO • How ARC-AGI measures the human-AI capability gap • Why open-ended environments matter for invention and scientific insight • Whether training on verifiable tasks (RLVR) transfers to broader problem-solving • How to measure creativity in AI systems SPEAKERS François Chollet, Co-Creator, ARC Prize; Creator, Keras & ARC-AGI Fred Sala, Chief Scientist, Snorkel AI CHAPTERS 00:00 Introduction 01:31 ARC-AGI and the human-AI capability gap 02:28 Invention and open-ended environments 04:12 Measuring open-ended work and scientific insight 08:50 General intelligence or a collection of skills? 12:21 Transfer learning and RLVR 15:20 Measuring creativity 17:56 Benchmarks for open-ended capabilities 20:38 Agent swarms and open scientific research 23:41 The next phase of open-ended benchmarking RESOURCES ARC-AGI benchmark: https://arcprize.org/arc-agi On the Measure of Intelligence: https://arxiv.org/abs/1911.01547 Snorkel AI Open Benchmarks Grants: https://benchmarks.snorkel.ai/