
Weird Little Model Behaviors

A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)
Topic 15 of 18
Published Oct 18, 2026 · Clip 1:01:48–1:03:27 (1:38)
Emmanuel shares two 'weird little things' from interpretability research. First, models have a separate 'gut check' mechanism that flags impossibility (e.g., noticing 4 + 5 = 23 must be wrong without doing the math) — learned hunches. Second, if you tell a model to do one task while thinking about another thing, interpretability shows intermediate representations for the other thing, a behavior with no obvious training reason.