Why is it harder for ChatGPT to write a half-decent essay than to solve an 80-year-old math problem that no human could crack? Isn’t that a weird state of affairs? Can AI be superhuman and yet have such limitations? We can start to answer these questions by deconstructing intelligence.

In A Brief History of Intelligence, Max Bennett distills 600 million years of brain evolution into five intelligence breakthroughs. Each breakthrough solves a problem the previous machinery couldn't: steering toward valence, learning from sparse reward, planning by simulation, modeling other minds, and transmitting thought through speech and writing.

In our next benkyoukai, Harold Godsoe will match the past three years of AI development against Bennett's book, discuss what the book gets right and wrong, and what it predicts from 2026 onward. Which intelligence breakthroughs could the next generation of models plausibly claim? Which failure modes do they imply? How does it all relate to safety research? Join us to find the answers to these questions and much more.

— Max Bennett, A Brief History of Intelligence (2023)