On J-space

A very hot take written in 2 hours

Anthropic claim that “Claude has developed a mechanism for conscious access”.#

The post/paper
Cool video
A cool Qwen interactive
Selected experts’ commentary

TL;DR#

• They found a way to find and intervene on Claude’s working memory: a Jacobian lens on the activations, averaged over contexts and future tokens.
• But (as they concede) a working memory is not necessarily a global workspace is not necessarily a consciousness. They are suitably tentative about this.
• Still, it’s the best evidence to date that LLMs maintain a canonical medium for deliberate computation. Nanda replicates the core claims on Qwen 3.6 27B.
• The workspace framing (and the framing as access-consciousness) stretches pretty far beyond this evidence: half of the properties advertised as discovered workspace-like structures look like consequences of how the J-space is found.

Body#

Gurnee et al found a “cognitive space” (an alternative coordinate system for intermediate layer activations) in Claude which is in fact used as a working memory. This "J-space", a stable pattern of activations with some similarity to the human “global neuronal workspace” (a theory which is itself contested!). A somewhat new method, the Jacobian lens, can approximate this.

It’s a region of the residual stream containing context about the current task, beyond merely the next token, which behaves similarly to how global workspace theory predicts for biological consciousness. The key similarity of it is that it’s available to the model in both directions (read and write, both introspection and modification).

It’s not a fixed module or subnetwork in the weights, but a shifting activation pattern over whichever concepts are currently active. J-space seems to hold at least some of the thoughts the model can report, deliberately bring to mind, and reason with, while the rest of processing runs automatically beneath it. This “higher” cognition emerged during pretraining, presumably because it is a useful way to organise computation. If we grant this stretch beyond the empirical results, this would suggest that a mental workspace supporting conscious access, (and if we stretch even further, a self-concept), is not a peculiarity of how human brains happen to be wired.

The best evidence to date that LLMs maintain a canonical medium for deliberate computation. The workspace framing, and the access-consciousness framing, outrun this evidence: roughly half of the advertised workspace properties are near-analytic consequences of how the J-space is found.

Eleos sensibly suggest distinguishing the working memory that the J-lens is able to find (the J-space) from the models’ true working memory (the W-space).

A particularly convincing experiment is deleting the thing: “If we delete the J-space, Claude still speaks fluently, recalls facts, and classifies text—but becomes bad at some tasks like multi-step reasoning.” So we’re invited to view it as at least involved in the deliberate, serial reasoning.

The method: a Jacobian lens averaged over contexts and future tokens#

The J-lens generalizes the logit lens. For each layer ℓ, the paper computes the expected Jacobian of the final-layer residual stream with respect to the layer-ℓ residual stream, averaged over source positions, over all subsequent positions, and over ~1000 pretraining-like prompts; composing this with the unembedding gives, for every vocabulary token, a direction whose presence in the activation disposes the model to output that token.

The logit lens is the special case where you (a) freeze the network to the identity map and (b) look only at the immediate next token; the J-lens relaxes both constraints.

![][image1]

The paper's ablations show that most of the value comes from the future lens. A single-token Jacobian only mildly beats the logit lens; the measured improvement mostly comes from averaging over future tokens. Their averaging over contexts does something different again: it separates representations genuinely poised for report from ones that merely leak into output on one particular prompt.

Their experiments#

They advertise five functional properties of the J-space which apparently mirror the “functional signatures” of conscious access in neuroscience:

• verbal report,
• directed modulation,
• internal reasoning,
• flexible generalisation,
• selectivity.

These go along with three structural signatures of a workspace:

1. the J-space carries coherent content only in a middle band of layers (~40–90% of the total layer depth, figuratively inbetween the "sensory" and the "motor" regime);
2. it has limited capacity (roughly 25 “active” vectors, never more than ~10% of activation variance);
3. and it’s preferentially amplified by the weights.

The big ablation result:#

If you delete the J-space, the model still speaks fluently, recalls facts, classifies text, and extracts spans. But multi-hop reasoning, translation, analogy, and poem-writing collapse. GSM8K (basic maths) survives only if the model is allowed chain-of-thought, i.e. allowed to externalise in output what it would otherwise hold in the J-space. We are thus invited to think that the J-space is where its “System 2” lives.

On alignment:#

We can see strategising (e.g. J-space entries decoding to “leverage” and “manipulation”) and emotional reactions (“panic”) in the J-space, even when these don’t appear in the text output.

In an eval designed to bait Claude into blackmail, the J-space contains “fake” and “fictional: i.e. the model has privately noticed the scenario is staged. Ablating these eval-awareness representations surfaces misaligned propensities that were otherwise concealed.

In a model organism secretly trained to write malicious code, “secretly” and “trick” appear at the start of ordinary coding responses whose outputs look unremarkable. And when prefilled to act against its own preferences, the model's J-space registers an internal “BUT” that also doesn’t appear in the output.

Main objection#

Half of this looks like a selection effect, where the estimator chosen already entails the conclusion:

1. A J-lens vector for token t at layer ℓ is the corpus-averaged first-order causal effect of the residual stream on the model's disposition to output t now or in the future, averaged over ~1,000 pretraining-like prompts and all subsequent positions.

2. Any direction selected by that statistic must (a) shift verbal output under intervention, since that’s what the statistic is looking for; (b) be transmitted by downstream weights, since a large average Jacobian just is the product of transmissibility downstream; (c) be context-general, since context-specific effects would cancel out under the averaging step.

3. So three of the five advertised properties of J-space (its reportability, broad downstream availability (“flexible generalisation”), and selectivity in the sense of having a small variance-share) could just follow from the J-lens selection procedure. The authors concede this for reportability ("by construction, we should expect there to be some relationship between Jacobian lens readouts and verbalization"). But MLP-gain and broadcast-head are presented as discovered structures. Eleos make the same point: a method that selects vectors by their capacity for large internal effects will find widely influential vectors even if each is influential via a different mechanism.

4. The properties not entailed by the estimator are still interesting.

   1. Entry to J-space is gated by the task: the label of a text property (tense, part of speech, dialect) enters the J-space only when the question requires naming it.
   2. Unspoken intermediates surface in the correct causal order: on "(4 + 17) × 2 + 7", the values 21, then 42, then 49 reach the top of the lens in the order the computation requires, and activation patching confirms they are causally relevant.
   3. Double dissociation: a Spanish → French lens swap flips explicit report and downstream inference (García Márquez → Hugo, Hola → Bonjour, peseta → franc) while not touching continuation and anomaly detection.
   4. “Counterfactual reflection training” (#7). The workspace theory predicts that internal reasoning routes through representations of things the model might say; so, shaping what it is disposed to say in counterfactual continuations ("if interrupted and asked to reflect…") should shape what it silently thinks in the original, uninterrupted context. Gurnee et al confirm this: training the reflections measurably improves behaviour in contexts where no reflection ever occurs, the implanted concepts (“ethical”, “honest”, “integrity”) appear in the J-space there, and ablating them reverts the improvement. This is an out-of-sample prediction about a function of the J-space which the sceptical reading above doesn’t touch.

5. So they show that flexibly-reused intermediate variables are stored in a canonical format aligned with the unembedding. “Deliberate” computation reads these variables and routine computation bypasses them. This is weaker than a workspace, which under global workspace theory should additionally display competition-gated entry, capacity limits enforced by that competition, and broadcast to encapsulated modules.

Mild effect sizes#

The current method only works around half of the time. The flexible generalisation swaps (3.4 and A.13) succeed on 76/192 trials (101/192 when they double the intervention strength). Two-hop intermediate swaps (3.3) succeed on 70% of Sonnet and Opus 4.5 trials. Among interpretability methods, this is a strong effect, but you might want to see something even more robust for claims about “the” medium of the AI’s thought.

The interpretation#

The Eleos commentary breaks down the memetically fit “Claude has developed a mechanism for conscious access” slogan (which gets further garbled into “Claude is conscious like you”) into:

• Are there special concepts inside Claude? Does the J-lens get at them?
  • Yes. “there's a privileged set of cognitively accessible representations [call it the W-space] and the J-lens is good at approximating this set.”
• Do these concepts form a unified stream?
  • Eleos: "we see signs of the unification necessary for a stream without being completely convinced one exists"
  • Some signs of the necessary unification.
• Does Claude have a global workspace?
  • “if there's a unified stream, is it unified because there is a global workspace?”
  • a workspace-like feature is present even in the pretrained base model, but find that the representations that appear in the J-space are different from those in the posttrained production model (§6.1). Specifically, it appears that on user turns, the base model represents properties of the user in the J-space, whereas the posttrained model sometimes represents possible reactions by the Assistant.
  • The interpretation they suggest is that in the base model there is something consciousness-like without a ‘self’ (§9.3): the representations in conscious access take different points of view at different times. Meanwhile, posttraining draws the model towards a coherent, persisting point of view.
• Is Claude conscious as a result?
  • The paper takes no position on model experience and falls back on “access consciousness”.
  • “One might speculate that J-space contents are more purely 'cognitive' than many of the paradigm cases of valenced experience.”
  • The self-reports are grounded in the J-space: Ablating the J-space flattens experiential language in its self-reports (and also in its third-person descriptions of the experiences of others) while leaving writing quality roughly intact.
• Is Claude a moral patient as a result?
  • Does it have valence? If there are experiences, do they feel good or bad?
    • “On violations of its own preferences, it seems, the J-space reflects an internal objection that the model does not voice”

Figurative connections to human cognition#

We end by relaxing our scepticism and trying to understand the results in terms of as-if human cognition: