François Chollet: How We Get To AGI

François Chollet discusses the pursuit of Artificial General Intelligence (AGI), contrasting the past focus on scaling up pre-training with the current shift towards test-time adaptation. He defines intelligence as the efficiency of using past information to handle novel situations and critiques exam-like benchmarks for measuring AGI progress, advocating for benchmarks that target fluid intelligence. Chollet introduces the Abstraction and Reasoning Corpus (ARC) benchmarks, including the new ARC AGI 2 and upcoming ARC AGI 3, designed to assess compositional generalization and agency. He proposes that true AGI requires combining two types of abstraction: value-centric (Type 1) and program-centric (Type 2), integrating deep learning with discrete program search. Chollet envisions AI systems that function like programmers, synthesizing software using a library of reusable abstractions, and highlights the work at his research lab, ENVIA, to build AI capable of independent invention and scientific discovery.