This is Francois Chollet’s keynote from the AGI-24 conference, recorded in 2024 and published in October that year. Chollet, the creator of the Keras deep-learning library and of the Abstraction and Reasoning Corpus, uses the talk to argue that simply scaling up large language models will not produce human-level general intelligence.
His central claim is that today’s models excel at pattern recognition over their training data but struggle with genuine abstraction and on-the-fly reasoning about novel problems. He uses ARC, a benchmark of puzzle-like tasks that humans solve easily but where models long scored poorly, to make the gap concrete, and he sketches a path that combines deep learning with program synthesis rather than relying on parameter count alone.
This is a useful counterpoint to the scaling-centric narrative that dominates AI discussion. For a general or business reader, it is a clear, firsthand explanation from a respected researcher of why benchmark progress on familiar tasks may not translate into the flexible reasoning that real general intelligence would require.