Inside the black box of AI: How AI models actually think | Emmanuel Ameisen (Anthropic)
- ID
- 36760
- Status
- summarized
- Published
- 18 Oct 2026, 8:31 PM
- Fetched
- 18 Oct 2026, 8:51 PM
- Provider
- Lenny's Newsletter
- Category
- product-startup
- Original URL
- https://www.lennysnewsletter.com/p/a-mind-expanding-conversation-on
- Source URL
- https://www.lennysnewsletter.com/feed
Summary
- Score
- 5.0
- Created
- 18 Oct 2026, 8:51 PM
- Tags
- Audience
- ai_ml_learnersai_agent_usersdevelopers
What happened
A 72-minute Lenny's Podcast episode (published Oct 18, 2026) with Emmanuel Ameisen, an interpretability researcher at Anthropic who previously worked in applied ML at Stripe and Insight and wrote Building Machine Learning Powered Applications. The conversation argues the 'next-token predictor' framing undersells how models plan many tokens ahead without stating those plans, that models are 'grown' rather than built, and that a 'J space' relates to model self-awareness. It ties interpretability work to alignment and safety, and references Anthropic's 'Tracing the thoughts of a large language model' and 'A global workspace in language models' write-ups plus the original Scaling Laws paper.
Why it matters
The concrete claim worth testing is that models plan tokens ahead without exposing those plans — which means judging an agent by its visible output alone can mislead you about why it took a step. If you debug agents or write evals, this episode is a prompt to check whether your failure analysis reads only final text; the linked Anthropic tracing/global-workspace research is the actual material to read, since the episode itself is paywalled and the excerpt contains no measurements, numbers, or reproducible findings. Nothing here is Malaysia-specific, so treat it as background rather than a local decision input.
Discussion angle
Ameisen says models plan many tokens ahead without stating those plans — if true, what would you change about how you debug a bad agent run: reading the transcript, or trying to inspect intermediate representations? Ask the room who has actually read the Anthropic tracing or global-workspace write-ups versus just the headlines.