Anthropic interpretability researcher: LLMs are 'grown' not built, and they plan many tokens ahead
lennysan · x · 2026-10-18
Lenny's Podcast hosts Emmanuel Ameisen, who has spent 3 years doing interpretability research at Anthropic, looking inside models like Claude to see what they're actually doing.
Key topics:
- His discovery that models plan many tokens ahead
- Why anthropomorphizing AI can actually be the right approach
- What the 'J-space' reveals about how models think
- How interpretability connects to AI alignment and safety
His framing: models are 'grown,' not built — 'to make them is not to understand them.'
More from AGI Musings
- Researcher echoes Mariya's critique: AI research is losing curiosity-driven work, offers open collaboration — mariyaivasileva · 2026-10-18
- 'Junior hiring isn't cooked' crowd mocked as WSJ advises job seekers to update LinkedIn — felpix_ · 2026-10-18
- Researcher Mariya Ivasileva warns AI research culture is eroding intellectual honesty — mariyaivasileva · 2026-10-18
- Princeton PhD applies to 2,300 jobs, still unemployed as junior-hiring debate rages — felpix_ · 2026-10-18
- MIT lets a Claude-powered agent freely "play" on a simulated island for 30 hours — and finds play expands its capabilities — LuizaJarovsky · 2026-10-18
- Using AI well demands more skill than most assume, argues researcher — davidmanheim · 2026-10-18