Summary
Anthropic researcher Jeremy Berman gives an intuitive, high-level tour of how large language models are built, in conversation with Matthew Berman. Berman is a self-taught AI researcher, and the session is aimed at anyone who wants a starting point for understanding the field.
The talk is organized into seven parts: what neural networks are, how they are trained, what language models are, pre-training, post-training, how models learn to think, and whether they can discover new things. It begins with a toy example of predicting New York apartment prices from location, square footage and bathrooms, showing how a network’s adjustable parameters, or knobs, are tuned using a loss function and gradients in PyTorch.
From there it moves to language models, which learn the same patterns as sequences of tokens, and on to the techniques behind systems like ChatGPT and Claude, including reinforcement learning and reasoning. Viewers get a clear mental model of the full training pipeline without needing a deep technical background, plus pointers for going further on each topic.
📺 Source: Matthew Berman · Published November 12, 2026
🏷️ Format: Interview
