Anthropic interpretability researcher: models plan rhymes before writing a line
lennysan · x · 2026-10-19
Lenny Rachitsky interviews Emmanuel Ameisen, who spent 3 years on interpretability research at Anthropic. Key points: LLMs are "grown, not built" — the only way to understand them is to look inside; models plan many tokens ahead (a Claude model internally chose "rabbit" as the rhyme before writing the line); anthropomorphizing AI can be the right approach; what "J-space" reveals about model thinking; how interpretability connects to alignment and safety.
Related event: Anthropic Researcher: LLMs Are Grown, Not Built, and Plan Ahead(3 posts)→
More from Companies & People
- Meta's Alexandr Wang endorses Meta over rival labs: 'highest use of superintelligence is invention' — alexandr_wang · 2026-10-19
- Hugging Face kicks off Open Source AI Week with ExecuTorch edge-AI hackathon — Thom_Wolf · 2026-10-19
- Why Oracle customers stay: the real answer is switching costs — richardreis · 2026-10-19
- Anthropic's coding AI is accelerating — yet you still can't change your Claude email without deleting your account — emollick · 2026-10-19
- Anthropic commits 5% of its shares to a 'public benefit fund' — TechCrunch AI · 2026-10-19
- Deep learning scientist Florentin Guth joining MIT EECS as assistant professor — m_e_sander · 2026-10-19