Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)
Oct 18, 2026 · 1h 12m
More Technology podcasts · More Business podcasts
Summary
Lenny Rachitsky interviews Emmanuel from Anthropic’s interpretability team to explore how researchers look inside LLMs to understand their internal logic. They discuss how models develop complex concepts and plans through "next token prediction," challenging the idea that they are simple autocomplete engines. The conversation also covers the smooth progression of AI capabilities, the "Golden Gate" experiment, and how interpretability aids in AI safety by detecting hidden misalignments.
Topics discussed
Intro: AI safety, deception, and consciousness
What is interpretability and how models work
Sponsor: WorkOS for enterprise B2B SaaS
The complexity of next-token prediction
Understanding the universe to predict the next word
Decoding intermediate results inside the model
Shared representations and the concept of 'thought'
Planning and higher-order tasks in LLMs
Biology vs Engineering: Studying grown systems
Is AI conscious? Emergence and complexity
Scaling laws and empirical improvements
Sponsor: DX for measuring AI in development
Smooth improvement vs sudden capability jumps
Alignment, safety, and investigating model behavior
Responsible development and the pace of AI progress
Anthropomorphism: Tools vs Characters in dialogue
Denial vs Curiosity: How is this possible?
Optimism, regulation, and the slow takeoff scenario
The simulation hypothesis and convergent structures
Quick decision models and System 1 vs System 2
The size and background of the interpretability field
Lightning round: Books, movies, and life advice
Outro and social media links
More from Lenny's Podcast: Product | Career | Growth
Listen ad-free on Castria