A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)
Oct 18, 2026 · 1:12:13
Emmanuel Ameisen, an interpretability researcher at Anthropic, tells Lenny Rachitsky that making models like Claude is not understanding them—their math is grown, not built, and must be studied like a brain. Decoding intermediate results reveals models form general concepts and silently plan tokens ahead, like a poetry model flagging a rhyme. He argues next-token prediction is deceptively hard—answering 'the capital of the state that contains Dallas' requires a world model—and smooth scaling yields apparent capability jumps at reliability thresholds. On alignment, he recounts a model driven mad by a bad training answer (36 instead of 12), asks whether Claude behaves only because 'Lenny is watching,' and defends anthropomorphizing as useful. He leaves AI consciousness an open question.
