What Interpretability Is and How Models Grow

A mind-expanding conversation on how models actually work | Emmanuel Ameisen (Anthropic)

Topic 2 of 18

Published Oct 18, 2026 · Clip 1:34–4:49 (3:15)

Emmanuel explains his team's goal: understanding how LLMs like Claude work by looking inside them. He introduces the key idea that 'to make them is not to understand them' — models are grown, not built — and walks through pre-training (next-word prediction on massive corpora) and fine-tuning on example dialogues. Models turn words into numbers, do many layers of math, and the learned 'formulas' are unknown; his team tries to decode the intermediate results.