AI and the World's Hard Problems · Theoretical computer science · Anthropic · 2026-10-13
Mapping the mind of a large language model
Researchers identified how millions of concepts are represented inside Claude Sonnet, a deployed production-grade language model, which they describe as the first detailed look inside a model of this kind. The model's internal state is a long list of neuron activations with no clear meaning, and each concept is spread across many neurons. Building on earlier work that used dictionary learning to match activation patterns to concepts, the team says this could help make AI models safer in future.