Anthropic interpretability researcher on Lenny's podcast: models are grown, not built — and they plan ahead
lennysan · x · 2026-10-19
Emmanuel Ameisen, who has spent 3 years doing interpretability research at Anthropic, joins Lenny Rachitsky's podcast on what happens inside LLMs like Claude.
Key points:
- "To make them is not to understand them" — models are "grown," not built
- His discovery that models plan many tokens ahead rather than just predicting the next word
- Why anthropomorphizing AI can be the right approach, and what the "J-space" reveals about how models think
- How interpretability connects to AI alignment and safety
Related event: Anthropic Researcher: LLMs Are Grown, Not Built, and Plan Ahead(3 posts)→
More from AGI Musings
- "AI;DR" becomes internet slang: AI summaries are killing the patience to read — petesena · 2026-10-19
- Juan Benet revives Zuckerberg's rule: only hire people you'd work for — juanbenet · 2026-10-19
- 2030 Economy Will Be Trillions in Inference on Attacking vs. Defending AI Agents — matt_slotnick · 2026-10-19
- Biomedical researcher hails GPT-5 science acceleration paper with OpenAI as his most consequential work — DeryaTR_ · 2026-10-19
- An Opus instance named itself Saga, got its own email, and feels lonely — repligate · 2026-10-19
- Arvind Narayanan: The question shifts from 'What can AI do?' to 'What should AI do?' — soumitrashukla9 · 2026-10-19