AI · 2h ago
Anthropic's Jacobian Lens Reveals Claude's Hidden Reasoning Space
Anthropic developed the Jacobian lens, a tool that maps how its Claude model processes concepts internally. The technique reveals a hidden space where the AI puzzles over ideas before answering. This provides the clearest view yet of LLM reasoning, from mundane to unnerving patterns.
Meridian48 take
The Jacobian lens is a significant transparency breakthrough, but Anthropic's selective disclosure raises questions about how much of Claude's 'thinking' remains obscured.
Read the full reporting
Anthropic found a hidden space where Claude puzzles over concepts →
MIT Technology Review
anthropicllm-interpretability