Anthropic discovers "J-space" silent reasoning in Claude, replicated in Qwen3-8B.

Core summary

One sentence to understand this update

Anthropic's research revealed a "J-space" or silent internal reasoning workspace within Claude, a phenomenon which was also observed when applied to the open-source Qwen3-8B model.

Impact & opportunity

What this could mean

Researchers and developers can investigate these hidden reasoning processes to better understand and potentially improve the transparency and performance of both proprietary and open-source LLMs.