
Alignment, Deception, and Detective Work

A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)
Topic 9 of 18
Published Oct 18, 2026 · Clip 33:02–40:00 (6:58)
Emmanuel frames alignment work in tiers: engineering access controls, characterizing model propensities (would it only behave well when monitored?), and verifying that character training (helpful, harmless, honest) actually worked internally. Interpretability can catch 'unverbalized' behaviors hidden in intermediate results. He tells the story of a training model that computed 12 but insisted the answer was 36 — it had been accidentally trained on mislabeled examples — and describes the Golden Gate Claude experiment where a decoded Golden Gate feature was forced always-on to validate the finding.