Article Source

• François Chollet on OpenAI o-models and ARC

François Chollet on OpenAI o-models and ARC

Abstract

François Chollet discusses the outcomes of the ARC-AGI (Abstraction and Reasoning Corpus) Prize competition in 2024, where accuracy rose from 33% to 55.5% on a private evaluation set. They explore two core solution paradigms—program synthesis (induction) and direct prediction (“transduction”)—and how successful solutions combine both. Chollet emphasizes that human-like reasoning requires both fuzzy pattern matching (deep learning) and discrete, step-by-step symbolic processes. He also reveals his departure from Google to establish a new research lab focused on program synthesis, and provides insights into the next-generation ARC-2 benchmark.