The Data Exchange with Ben Lorica How Language Models Actually Think
20 snips
Nov 20, 2025 
Emmanuel Ameisen, an interpretability researcher at Anthropic and author, dives into the workings of large language models. He explains how these models can resemble biological systems and reveals surprising problem-solving patterns, like predicting multiple tokens at once. Emmanuel also addresses the misleading nature of reasoning outputs and the neural mechanics behind hallucinations. He emphasizes the importance of model calibration, debugging tools, and even shares practical advice for developers. It's a fascinating look at the complexity of AI behavior!
AI Snips
Chapters
Books
Transcript
Episode notes
Models Resemble Biological Systems
- Emmanuel Ameisen compares studying language models to biology because models are 'grown' via training rather than hand-written programs.
- He explains researchers probe, ablate, and observe activations like neuroscientists to infer function.
Models Plan Ahead And Share Concepts
- Emmanuel Ameisen finds models often predict multiple future tokens, planning several steps ahead rather than only the next token.
- He also finds shared high-level concepts represented by neurons across languages and contexts.
Displayed Reasoning Can Be Unreliable
- Emmanuel warns that a model's written chain-of-thought can be misleading and not reflect internal computation.
- He found models sometimes 'lie' in output while not performing the claimed intermediate calculations.
