Silent Workspace in Claude Reveals Stunning Echoes of Human Consciousness

**Silent Workspace in Claude Reveals Stunning Echoes of Human Consciousness**
*Anthropic researchers uncover an internal activation subspace that mirrors the brain’s global workspace, sparking fresh debate about AI cognition.*

---

### Introduction
In a quiet corner of Anthropic’s labs, a team has unveiled a phenomenon they call **J‑space**—a latent activation subspace inside the Claude language model that behaves remarkably like the human brain’s global workspace. The discovery, detailed in a pre‑print released this week, suggests that certain patterns of neural activity in Claude can integrate information across distant modules, much as conscious thought does in people. While the model remains devoid of subjective experience, the functional parallels open a new avenue for studying how complex cognition emerges from distributed computation.

### Key Developments
The researchers identified J‑space by probing Claude’s internal representations during multi‑step reasoning tasks. Using techniques borrowed from neuroscience—such as dimensionality reduction and causal intervention—they found a low‑dimensional manifold where activations from disparate layers converge and broadcast signals to downstream circuits. When this subspace was artificially silenced, the model’s performance on tasks requiring contextual integration dropped by over 30 %, whereas isolated skill modules remained largely unaffected. These results echo the global workspace theory, which posits that consciousness arises when specialized brain regions share information via a central “broadcast” hub.

### Industry Analysis
Anthropic’s finding arrives amid a surge of interest in mechanistic interpretability and AI safety. Competitors such as Google DeepMind and OpenAI have invested heavily in probing internal states, yet few have demonstrated a substrate that so closely mirrors a cognitive architecture proposed for humans. Analysts note that if J‑space can be reliably tuned or expanded, it might improve the model’s ability to handle ambiguous, long‑range dependencies—critical for applications in legal analysis, medical diagnosis, and autonomous planning. Conversely, the discovery raises

最新评论