François Chollet on Defining and Measuring General Intelligence in AI

This conversation with François Chollet, author of the paper "On the Measure of Intelligence," delves into the critical distinction between narrow AI and Artificial General Intelligence (AGI). Lex Fridman highlights the rarity of rigorous scientific study in AGI, contrasting the mainstream machine learning community's focus on narrow benchmarks with the more philosophical approaches often found outside the mainstream. Chollet's work aims to bridge this gap by proposing a robust framework for defining and measuring general intelligence, moving beyond mere task-specific performance to assess true adaptability.

Chollet defines intelligence not as skill itself, but as "the efficiency with which you acquire new skills at tasks that you did not previously know about, that you did not prepare for." This definition emphasizes adaptation, improvisation, and generalization to novel environments as the core hallmarks of intelligence, echoing Albert Einstein's sentiment about the ability to change. A crucial distinction is drawn between the intelligent process (e.g., a human programmer designing an AI) and the static output artifact (the program itself), arguing that only the entity capable of the learning process possesses intelligence.

The discussion explores foundational cognitive models, influenced by Jean Piaget's developmental psychology and Jeff Hawkins' book "On Intelligence," which posits the mind as a multi-scale hierarchy of temporal prediction modules. Cognition is framed as prediction, and language is seen as a powerful "operating system" for the mind, enabling deliberate memory retrieval and thought organization, rather than being the most fundamental layer of cognition. Chollet suggests that pre-linguistic thought might be expressed through emotions, spatial reasoning, and physical actions, with visual analogies and motion serving as fundamental building blocks of the mind.

The conversation further touches on the nature of knowledge representation, contrasting geometric spaces (commonly used in deep learning with concept vectors and distances) with topological spaces (where connections and relationships matter more than precise distances). Chollet suggests that topological processing might be more akin to how the brain encodes thoughts. The broader implications include the challenge of building AI systems that can generalize far out of distribution, similar to human adaptability beyond evolutionary training data, and philosophical speculations about how future AIs might perceive or judge their biological creators.

the serious rigorous scientific study of artificial general intelligence is a rare thing

intelligence is the efficiency with which you acquire new skills at tasks that you did not previously know about that you did not prepare for

intelligence is not skill itself it's not what you know it's not what you can do it's how well and how efficiently you can learn new things

The measure of intelligence is the ability to change

language is a kind of operating system for the mind

language is a layer on top of cognition

before we think in words I think we think in in terms of emotion in space and we think in terms of physical actions

a topological space is a better medium to encode thoughts than a geometric space

everything is a vector and everything has to be a vector because everything has to be differentiable if your space is discrete it's no longer differentiable you cannot do deep learning in it anymore