Anthropic's J-Space lens reveals AI models' internal Global Workspace

Anthropic has released groundbreaking research that peers into the hidden reasoning of large language models. Their new J-Space lens technique reveals a 'global workspace'—an internal representation the model uses while deciding what to generate. Unlike traditional methods that only see final outputs, J-Space exposes the dynamic, intermediate thought process, showing how the model weighs options, corrects itself, and prioritises concepts before producing text. The release includes the complete J-Space lens code, allowing developers and researchers to apply this interpretability tool to any compatible model.

To demonstrate the power of J-Space, an Anthropic partner created a live demo on Qwen 3.6 27B, a large open-source model. The visualisation tracks the model's internal state in real time as it works through prompts, revealing shifts in attention, confidence, and conceptual linkages. This marks a major step toward understanding AI 'thinking'—moving beyond black-box outputs to genuine transparency. For safety researchers, it offers a new way to detect biases, hallucinations, or hidden goals. For developers, it promises deeper debugging capabilities and more trustworthy AI systems.