# J-Space

> **TL;DR.** J-Space is the small, sparse subset of a language model's internal activations that Anthropic researchers found behaves like a conscious workspace: a privileged zone available for verbal report, top-down control, and multi-step reasoning, sitting atop a far larger volume of automatic computation.

- **Category:** AI / Interpretability / Model Safety
- **Stage:** nascent
- **Age:** 2 days
- **Origin date:** 2026-07-06
- **First detected:** 2026-07-08
- **Canonical URL:** https://earlyterms.com/term/j-space
- **Sources:** 6 primary URLs

## Definition

J-Space is the small, sparse subset of a language model's internal activations that Anthropic researchers found behaves like a conscious workspace: a privileged zone available for verbal report, top-down control, and multi-step reasoning, sitting atop a far larger volume of automatic computation.

Anthropic's Transformer Circuits team, led by [Wes Gurnee and Jack Lindsey](https://transformer-circuits.pub/2026/workspace/index.html), published the finding July 6, 2026 in "Verbalizable Representations Form a Global Workspace in Language Models," introducing the Jacobian lens (J-lens) that surfaces which concepts Claude is internally poised to verbalize at any given step.

## Example

In one demo, Anthropic told Claude to think about the Golden Gate Bridge while it copied an unrelated sentence; the J-lens showed "bridge" and "California" active in J-Space even though neither word appeared in Claude's visible output, per [Anthropic's accompanying video](https://x.com/AnthropicAI/status/2074185358678364414).

## Analogy

J-Space is the lit stage in a theater of dark backstage machinery — only what's on stage gets narrated.

## Why it's emerging now

Anthropic published "Verbalizable Representations Form a Global Workspace in Language Models" on July 6, 2026, claiming Claude models develop an internal workspace mirroring Global Workspace Theory of consciousness — and showing the same lens can flag hidden sabotage-related reasoning for safety auditing.

## Related terms

- *related:* Silent Sabotage Mode
- *related:* Claude Sonnet 5
- *related:* Claude Opus 4.8
- *related:* Mythos-class
- *child:* J-lens
- *parent:* mechanistic interpretability
- *related:* sparse autoencoders
- *parent:* global workspace theory
- *related:* golden gate claude
- *competitor:* chain-of-thought reasoning

## Sources

1. [Anthropic / Transformer Circuits — Verbalizable Representations Form a Global Workspace in Language Models](https://transformer-circuits.pub/2026/workspace/index.html)
2. [Axios — Anthropic says Claude has carved out its own space to ponder](https://www.axios.com/2026/07/06/anthropic-claude-ai-conscious)
3. [VentureBeat — Anthropic's new J-lens reveals a silent workspace inside Claude](https://venturebeat.com/technology/anthropics-new-j-lens-reveals-a-silent-workspace-inside-claude-that-mirrors-a-leading-theory-of-consciousness)
4. [The Unfinishable Map — Research notes on Anthropic's Global Workspace / J-Space](https://unfinishablemap.org/research/anthropic-global-workspace-j-space-2026-07-07/)
5. [Hacker News — J-Space: Where Claude silently performs reasoning steps](https://news.ycombinator.com/item?id=48825315)
6. [Anthropic (@AnthropicAI) — J-Space sabotage-detection thread](https://x.com/AnthropicAI/status/2074185373341688258)

---
_Generated by EarlyTerms · https://earlyterms.com/term/j-space_