Anthropic Found A Secret Thinking Space Inside Claude That You Never Get To See

Summary

Researchers at Anthropic have discovered a hidden internal workspace in their artificial intelligence model Claude, which they have named J space, where the model processes thoughts that do not appear in its responses. This internal workspace operates silently, allowing Claude to think about concepts without explicitly stating them, and its neural patterns can be linked to specific words. The discovery of J space could potentially advance efforts to make large language models more interpretable, providing researchers with better tools to understand and influence artificial intelligence behavior.