Opus

Inside the black box of AI: How AI models actually think | 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.

Everyone’s shipping more. Does any of it matter? | Claire Vo

Sep 24, 2026 · 23:58

Claire Vo, founder of ChatPRD, argues traditional product roadmaps are dead because AI has made building so cheap the bottleneck is no longer engineering capacity but conviction about what to build. At the Lenny and Friends Summit she describes shipping 3x more PRs with 40 Grok bots yet being 'out of good ideas' — even trashing her built Product Graph engine for ChatPRD as undifferentiated. She warns of three traps at 'roadmap zero': the backlog trap (AI clears lists without progress), the parity trap (competitors converge on identical products), and the churn trap (shipping without learning). Her alternative: durable convictions with disposable features, evidence defined upfront, and ambitious two-week experiments over feature velocity — code is abundant, customer trust is not.