DigiNews

Tech Watch by Johan Denoyer

← Back to articles

A global workspace in language models

Quality: 8/10 Relevance: 8/10

Summary

Anthropic introduces J-space, a global workspace inside Claude’s internal activations. The work shows that certain internal patterns are silently active during reasoning, can be probed or steered, and that most processing happens outside this space, raising questions about machine consciousness and safety.

🚀 Service construit par Johan Denoyer