Lenny's Podcast: Latest Episode
Newest Lenny's Podcast episode: A mind-expanding conversation on how models... (October 18, 2026). Key takeaway, summary and previous episodes from 1% Better.
Page updated by 1% Better. New episodes are summarised within hours of release.
A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)
Lenny's Podcast · Published · 1h 12m
Key takeaway
Pay attention to what AI models do between the prompt and the answer, not just the answer itself. Next time you use a tool like Claude, notice that it is not simply retrieving a memorized response. According to Emmanuel's research, the model forms intermediate representations of concepts, builds plans, and sometimes catches its own errors before producing a word. Emmanuel's team finds that the models represent ideas the way humans do, with shared internal patterns for the same concept across different contexts. Understanding this helps you use AI more thoughtfully and set more realistic expectations for what it can and cannot do.
Episode overview
How do AI researchers actually study what happens inside a language model? And what have they found that surprises even the people who built it? Emmanuel, an interpretability researcher at Anthropic, studies Claude by examining the intermediate numerical states it produces while generating text, the same way a neuroscientist scans a brain to find which regions activate for which stimuli. The conversation covers three main areas: - **How models are made vs. how they are understood.** Training produces a system that no one fully comprehends, and a separate research effort is required to decode what it learned. - **What interpretability has revealed.** Models plan ahead, hold abstract concepts internally, and can represent things they never say out loud. - **AI safety and alignment.** Interpretability is one tool for checking whether a model's trained values actually match its internal states. Interpretability refers to the field of research that tries to decode the internal computations of AI models rather than just observing their outputs.
Read the full Lenny's Podcast episode summary for key insights, quotes and action items.
Previous Lenny's Podcast episodes
- Why every company now needs to think and operate like a lab team | Josh Woodward (VP Google Labs)
- Molly Graham: The grief, burnout, and opportunity hiding inside the AI transition
- 90 minutes of unfiltered product advice from Snap and Discord’s product chief | Peter Sellis
- AI’s third era: the rise of persistent AI coworkers | Tara Seshan (OpenAI’s product lead)
- 84 minutes of enterprise sales alpha | Jen Abel
About this page
This page always shows the newest Lenny's Podcast episode in the 1% Better library. We summarise each episode from its audio or captions within hours of release, so bookmark it to check what is new. Browse every Lenny's Podcast summary and guide or all episode summaries.