Concepts, Features, and Hidden Planning

A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)

Topic 4 of 18

Published Oct 18, 2026 · Clip 11:50–18:33 (6:43)

Emmanuel describes how his team looks inside models at intermediate results and finds models represent concepts like 'Texas' and 'capitals' and combine them. Concepts are represented uniquely across contexts (an eye in a photo, diagram, or drawing activates the same feature). He then explains the big finding that models plan: when writing poetry, intermediate results encode a plan to rhyme (e.g., with 'rabbit') that future words refer back to, even without saying the plan out loud. Planning and next-token prediction are compatible because predicting the next instant well requires understanding the future.