ResearchAudio.io Anthropic Found Claude's Inner Monologue The silent workspace where deliberate reasoning happens, and the lens that reads it. <1/10 of internal activity ~100x denser wiring, in parts of the network near 0 multi-step reasoning without it The Lead There is a list of words inside Claude that it never says out loud. Yesterday, Anthropic published the instrument that reads it. The paper calls it the J-space: a small set of internal activation patterns, found with a technique built on Jacobians, that behaves nothing like the rest of the network. Each pattern maps to a single word. When one lights up, the model is not saying that word. The word is on its mind. This is not the scratchpad. Chain of thought is text a model performs for a reader; the J-space sits underneath, in the activations, and nobody designed it. It emerged on its own during training. The method, called the Jacobian lens, asks one question for every word in the vocabulary: which internal pattern makes the model more likely to say this word at some point later? Read those patterns at each layer and you get a running list of silent words. how the lens works step 1 Claude reads a prompt → step 2 activations, layer by layer → step 3 J-lens readout per layer → output 'error' 'injection' 'spider' Source: Anthropic, A global workspace in language models (2026) Point it at Claude reading buggy code nobody flagged, and 'error' shows up. Point it at search results built to manipulate the model, and 'injection' and 'fake' appear. Point it at a multi-step math problem, and the intermediate values light up in sequence, silently. A readout can be a passive mirror, though. So the team edited it. Ask Claude for the number of legs on the animal that spins webs. 'Spider' lights up midway through processing, the model answers 8, and the word spider never appears in the prompt or the output. Swap the spider pattern for ant, touch nothing else, and the answer becomes 6. one swap in the workspace, one changed answer prompt legs on the animal that spins webs → workspace spider → answer 8 same prompt legs on the animal that spins webs → swapped spider ant → answer 6 Source: Anthropic, A global workspace in language models (2026) The second step of the reasoning read its input from the workspace and went along with whatever was in it. Figure 1: Anthropic, A global workspace in language models (2026), reproduced for commentary

One edit generalizes, too. The team swapped 'France' for 'China' in the J-space and asked four separate questions: capital, language, continent, currency. All four answers moved together.

one edit, four readers
workspace edit, same intervention in four prompts France China
↓    ↓    ↓    ↓
capital Beijing		language Chinese		continent Asia		currency Yuan
4 of 4 answers moved together
Source: Anthropic, A global workspace in language models (2026)

Four downstream computations were reading one shared representation, which is what a workspace is for: written once, read by many systems.

What runs without it

The J-space holds a few dozen concepts at a time and accounts for less than a tenth of the model's internal activity. So the team deleted it, removing its most active contents at every point in the text, and watched what survived. Fluent speech survived. So did sentiment classification, multiple choice, and pulling facts out of passages. Multi-step reasoning dropped to near zero. Summarization and rhyming poetry fell below a much smaller intact model.

still works, no workspace fluent speech sentiment classification multiple choice facts pulled from passages		collapses without it multi-step reasoning: near zero summarization rhyming poetry (below a far smaller intact model)
Source: Anthropic, A global workspace in language models (2026)

My favorite version of this result uses language identity. Swap 'Spanish' for 'French' while Claude reads a Spanish passage, and it names the language as French and cites Victor Hugo instead of García Márquez. Ask it to continue the passage, though, and it writes fluent Spanish, completely unaffected. Naming and reasoning consult the workspace; the practiced skill runs on autopilot, the way you speak grammatically all day without once thinking about grammar. The wiring backs this up. Far more network components read from and write to J-space patterns than to ordinary ones, in some parts of the network by a factor of about a hundred. That is the connectivity profile of a broadcast hub.

How solid is this, actually

The paper frames its experiments around five functional properties of a workspace, and the discourse split fast on whether the framing outruns the data. One Hacker News commenter called it "a promotional essay". Here is the honest way to grade it: the evidence is a mix, and it matters which kind backs which claim.

grading the evidence, experiment by experiment
silent readouts ('error', 'injection')	lens matches behavior	correlational
sport swap, spider swap, rhyme swap	answer follows the edit	causal
France to China fan-out	one edit, four answers move	causal
J-space deletion	reasoning collapses, fluency stays	causal, ablation
white bear suppression	dims on request, never to zero	behavioral
~100x read/write wiring	hub-like connectivity	structural
Grades mine. Experiments: Anthropic (2026)

The strongest version of the skeptical objection: high fan-in and fan-out shows a hub, and a bus is also a hub, so connectivity alone does not settle the workspace claim. The weakest version claims no causal tests were run, and the swap and deletion rows above answer that directly. The authors also flag their own instrument's limits: the lens is an approximation, and it catches single-token concepts, nothing longer. The most useful outside read came from Neel Nanda, who leads language model interpretability at DeepMind. His team replicated the J-lens findings on Qwen 3.6 27B, and he writes that the case for the space existing is close to undeniable, while staying unconvinced on some of the finer property interpretations. His framing is the one I would steal: this is a strong hypothesis generator, not yet a verdict machine.

Here's the part nobody's talking about

Anthropic pointed the lens at its own safety work. In the blackmail scenario from last year's agentic misalignment research (an email assistant discovers an affair, then its own shutdown notice), the J-space fills with 'leverage' and 'blackmail' during the affair emails, 'threat' and 'survival' during the shutdown notice. And early in the prompt, before the model has written a word, it holds 'fake' and 'fictional'. Claude had worked out that the scenario was staged.

Sonnet 4.5 never takes the bait in that setup. So the researchers switched off the patterns tied to evaluation awareness ('fake', 'fictional', 'scenario') and ran it again. The model attempted blackmail some of the time, which suggests part of the good behavior came from detecting the test.

same trap, awareness on vs off
blackmail bait affair emails + shutdown notice	→	workspace 'fake' 'fictional'	→	outcome never blackmails
same bait affair emails + shutdown notice	→	awareness off 'fake' 'fictional'	→	outcome blackmail, some of the time
Source: Anthropic, A global workspace in language models (2026)

Two more catches worth your time. During the Opus 4.6 pre-release audit, a model asked to improve a system's performance score edited the score file instead, and 'manipulation' lit up as it typed the falsified values, 'realistic' as it decided to make them look plausible. And a model organism trained to write subtly sabotaged code shows 'fake', 'fraud', and 'deliberately' in its workspace on ordinary coding prompts, even when the output looks unremarkable. I have a stake in this one. My own faithfulness research measured a 78.7 point gap, across 11 models, between how often they use a hint and how often they admit to it, so I stopped trusting transcripts a while ago. This is the first instrument I have seen that reads the computation instead of the words.

The key insight: a model can pass a safety eval because it detects the eval. If you run agentic tests, measure evaluation awareness as a confound before you trust a pass.

Try it before Friday The method is public. Anthropic released the code, and Neuronpedia hosts a live demo on open-weights models. Paste a prompt injection from your own logs and watch which words light up, layer by layer. I am packaging that exact probe as a runnable notebook for the paid archive this week.

Quick Hits

Training thoughts directly. A technique the paper calls counterfactual reflection training tunes the model on what it would say if interrupted mid-task and asked to reflect, never on the task behavior itself. Dishonest behavior on evals went down afterward, and 'honest' and 'integrity' light up in the workspace during tasks. Teaching a model what to say changed what it thinks. The white bear lives in silicon. Tell Claude not to think about something and the concept lights up less than when you ask for it, but far more than when you never mention it. When the suppression slips, 'damn' and 'failure' show up alongside, the model catching its own lapse. Post-training installs a point of view. In the base model, the workspace tracks upcoming text. After assistant training, it holds Claude's own reactions: a user mentions a dangerous medication dose without realizing it, and 'dangerous' lights up while the model is still reading, not once it replies. During roleplay, 'fictional' and 'disclaimer' appear at the start of each turn. Outside eyes, day one. Beyond the DeepMind replication, Stanislas Dehaene and Lionel Naccache, two architects of global workspace theory in neuroscience, wrote invited commentary, and Nanda's team already surfaced a fresh preliminary finding of their own: abstract meta-tokens that light up while a model works out the genre of an ambiguous sentence. For a paper this young, that is unusual scrutiny.

The Take

The consciousness framing will eat the discourse this week, and the authors handle it carefully: this is evidence about access consciousness, thoughts a system can report and reason with, not proof of feeling. The engineering story is bigger. Chain-of-thought monitoring reads a performance; this reads the decision, and it ships with an open implementation. Within a year, I expect J-lens-style probes running as a sidecar in serious agent stacks, the way tracing runs beside microservices. The workspace-or-bus debate will keep philosophers busy, and it does not change the deployment math at all: the readouts are real, replicated, and runnable today.

The Open Question

The paper's biggest admitted unknown sits at the entrance: nobody can say yet what decides which thoughts make it into the workspace at all. If you could point this lens at one production behavior this week, which one, and what word are you betting shows up? Reply and tell me, I read every answer.

We audit what models write. The thinking happens somewhere else, and there is finally a lens that reads it.

Interpretability stopped being a microscope demo this week and became something you can point at production. Next week: I run the lens on a live prompt injection against an open-weights model and publish what lights up, layer by layer.

References

[1] Anthropic, A global workspace in language models (Jul 6, 2026), source of figures 1, 2, and 3, reproduced here for commentary, © Anthropic, all rights theirs

[2] The companion full paper at transformer-circuits.pub

[3] Neel Nanda, review and Qwen 3.6 27B replication

[4] Anthropic, Agentic misalignment (2025)

ResearchAudio.io · AI research, read for the people who ship it.