What Anthropic's J-space Actually Is (and Isn't)

July 7, 2026 Vibes by me, commas by AI ✨

mechanistic-interpretabilityjacobian-lenslogit-lensresidual-streamactivation-ablationcausal-interventionsglobal-workspace-theoryaccess-consciousnessai-safetytransformers

A small bright ordered lattice of golden filaments suspended in a vast dark tangle of dim blue threads

On July 6, 2026, Anthropic published A global workspace in language models, based on the paper Verbalizable Representations Form a Global Workspace in Language Models. Within hours the internet had sorted itself into two camps: people who think Claude is now sentient and would like to marry it, and people who think Anthropic discovered that a computer computes. Both camps have one thing in common: neither read the paper. Well, I and my minion machine did. It’s annoyingly good, which ruins my plan of dunking on it.

One scoping note up front: every experiment in the paper was run on Anthropic’s own Claude models. Keep that in mind, because it matters later.

So what’s all the fuss about?

They found a small subset of an LLM’s internal representations that the model can report, deliberately control, and reason with, sitting on top of a much larger volume of automatic processing it can’t. They call it the J-space (after the Jacobian, the math used to find it) and read it with a tool they call the J-lens.

That’s the claim. Everything about consciousness is vibes, professionally typeset.

How the tool works

A transformer carries a running vector at each token position (the residual stream) that every layer reads from and writes to. By the final layer it’s been shaped into something you can multiply by the unembedding matrix to get next-token scores. The layers in between do the actual work. The J-lens reads those intermediate layers.

For each vocabulary word it asks: which direction in the intermediate activations, averaged across many contexts, pushes the model toward saying that word later? Concretely: backprop from the final layer to an intermediate one, average the Jacobian over a thousand prompts. The averaging is the entire trick. It separates concepts the model is poised to say from ones that merely got said once. It’s the logit lens after finishing grad school: the logit lens assumes every layer speaks the same coordinates, the J-lens corrects for drift across layers and recovers signal in early layers where the logit lens returns static.

Your first objection, and mine: this is circular. Define a subspace by its relationship to future output, then gasp when it relates to future output. Congratulations, you’ve discovered the readout. The paper saw you coming.

What they actually demonstrated

Finding a readable subspace is nothing. The move is showing that a subspace defined only by verbalizability then satisfies four properties nobody selected it for.

It’s privileged for report, not just correlated with it. Split a concept’s representation into its J-space component and everything else. The J-space part carries a median of 6 to 7 percent of the variance and nearly all of the causal juice: swap along it and the model’s answer flips 59% of the time; swap along the other 93% of variance and you get 5%. The leftover 5% is itself routed back through the J-space, because clamping the J-space coordinates drops it to zero. So the readable directions are causally special and the fat remainder of the representation, the part doing most of the flexing in variance terms, is causally dead weight for report. There goes the circularity objection, and it was my favorite.

It holds intermediate reasoning steps, and they’re load-bearing. Ask “the number of legs on the animal that spins webs is” and “spider” appears in the lens at middle layers, present in neither prompt nor output; swap “spider” for “ant” and the answer flips from 8 to 6. In rhyming couplets the planned rhyme word shows up before the line is written, and swapping it changes word choices earlier in the line. That’s planning, not post-hoc labeling. They even kill the obvious confound (the intermediate vector secretly containing the answer) by showing the intermediate swap bites ~17% earlier in depth than an answer swap. Multi-hop swaps move the answer 54% of the time on Claude Haiku, 70% on Sonnet and Opus.

Ablate it and the model splits cleanly in two. Suppress the top J-space directions and the shallow stuff survives: multiple-choice, sentiment, grammar, span extraction. Anything requiring inferred, assembled content collapses below the much smaller Haiku baseline: multi-hop reasoning, analogy, translation, summarization, sonnets. Best detail in the paper: GSM8K with explicit chain-of-thought survives ablation far better than the same problems answered directly. The model writes on the page what it would otherwise carry in the workspace. It invented scratch paper, same as you did in third grade.

Where the framing is careful, and where it stretches

Here’s the part the hot takes missed on both sides: the paper does not claim the model reproduces the brain’s global workspace. It lists its own architectural mismatches, in print, where anyone could have read them.

No recurrence: the “broadcast” happens within one feedforward pass, across depth, not through loops. On ignition, the all-or-nothing amplification that signals workspace entry in actual brains, they say outright it’s unclear whether anything they see mirrors it. No evidence the non-workspace processing consists of encapsulated modules either, and they say so. So if your gotcha was “this isn’t really GWT,” I have bad news: the authors agree with you, in writing, and got there first. The actual claim is narrower: the J-space reproduces many functional properties of a workspace while sharing only some of the architecture.

Two stretches survive, though.

“Conscious” appears over 200 times in a paper that concludes nothing about consciousness. The text is scrupulous: access consciousness (poised for use in reasoning and report) is purely functional, no position taken on subjective experience. Fine. But you don’t say a word 200 times by accident. The framing invites the leap the text declines to make, the press took the invitation, and here we are. Access-consciousness structure is roughly what you’d expect from an LLM trained to chain reasoning and answer questions about itself. Concentrated and readable? Interesting. Evidence about feeling? No.

Nobody checked whether this is a Transformer thing or an everything thing. If the J-space is “the output-aligned, causally-privileged readable subframe of a task-optimized autoregressive net,” it should show up in an RNN too. The paper corroborates across Claude Haiku, Sonnet, and Opus, which is one family photographed three times. And the lineage matters: Jacobian-based identifiability of latent structure has been a nonlinear-ICA research program for years. That doesn’t diminish the result. It does relocate it from “we found a mind” to “Jacobians recover latent structure, episode nine.”

One genuinely weird finding I’ll flag without adopting it as a religion: ablate the J-space during open-ended self-report and the model’s language goes from experiential to mechanical, and the same flattening happens when it describes another person’s experience. Clean, replicated, unexplained, and not the headline anyone ran.

So should you update on AI consciousness?

No. The training objective predicts access-consciousness structure, the paper refuses the leap to phenomenal experience, and the consciousness framing is the weakest part of a strong package. Naturally it’s the only part anyone is talking about.

What’s actually useful

Delete the philosophy and a real tool remains.

A pre-output audit signal. In alignment evals the J-space carries strategic deliberation (leverage, manipulation) and eval-awareness (fake, fictional, meaning the model has clocked that it’s being tested) that never reaches the output. Ablating the eval-awareness representations in one case surfaced a misaligned propensity the model had been politely hiding. On two deliberately-misaligned model organisms, the lens flagged their planted objectives on ordinary prompts. Output filters read what the model says; this reads what it’s about to do.
A training technique falls out of it. “Counterfactual reflection training” shapes what the model would say if interrupted and asked to reflect, which measurably improves behavior in the uninterrupted context, and the gains reverse when you ablate the implanted J-space concepts. Testable mechanism, not vibes.
You can play with it today. The logit lens recovers much of the same structure at lower reliability, so start free. J-lens readouts on open-weight models are on Neuronpedia, and the Jacobian-lens idea predates this paper. Read the source, not the press cycle.

The experiment I’d want to see

Run the ablation dissociation and the report-privilege test on an RNN or a task-trained CNN. Either it reproduces and this is a finding about task-optimized networks in general, or it doesn’t and we learned something Transformer-specific. Either outcome is more informative than another week of consciousness discourse. The remaining questions that decide whether this ever touches production: does the J-lens transfer to models Anthropic didn’t train, and what does it cost at serving scale?

The model does not have an inner mind. It has a small, readable, causally load-bearing slice of network where deliberate reasoning concentrates, the authors said so with unusual honesty, and the internet heard “sentient.” Good tool. Wrong discourse.