Anthropic finds evidence of global workspace in Claude's reasoning

Anthropic researchers discovered that Claude appears to use a shared computational space—similar to human consciousness—when solving complex problems. The finding suggests language models concentrate information in specific layers before distributing it, which could improve how developers build reasoning-heavy applications and debug model behavior.