Anthropic’s J-Space: A Workspace Made of Signs. Why Vygotsky Explains the Data Better Than Baars

On July 6, 2026, researchers at Anthropic published the results of a new study claiming to have discovered an analogue of the «global workspace» in Claude, dubbed J-space.

The research is highly fascinating and, I believe, reveals the future of neural networks much more deeply than it appears at first glance.

In this article, I will explore why the Global Workspace Theory (GWT) accurately describes the structure of this finding, yet remains silent on its main oddity: the fact that this workspace consists of «words» (or, more precisely, discrete signs). And I will explain why Lev Vygotsky answered this exact question a century ago.

The Essence of the Research

In the course of their experiments, the researchers discovered a subcomponent of the model’s representational space (the J-space) that meets the following criteria (direct quote from the article):

• Verbal report. When the model is asked what it is thinking about, it names concepts represented in the workspace. Swapping one active workspace vector for another changes its answer to match.

• Directed modulation. When instructed to hold a concept in mind, or perform mental calculations, the model is capable of activating and computing with workspace vectors, independent of its outputs. In addition, information that is not typically represented in the workspace can be pulled in when the task requires it.

• Internal reasoning. Workspace vectors can be used to represent the value of intermediate computations, when the model chains inferential steps or composes plans, and intervening on them is sufficient to redirect the conclusion.

• Flexible generalization. The same representation serves as a valid argument to many different downstream computations. In other words, a workspace vector lifted from one context and placed in another is correctly operated on by whatever function the new context supplies.

• Selectivity. The workspace comprises a small subset of the total representational content of the model’s activations. It is required for only a fraction of the model’s behavior, and in particular is not involved in pervasive, routine processing like text parsing or grammatical fluency.

I won’t dwell on the technical details of the experiments; those interested can carefully review them by following the link to the original paper.

The researchers at Anthropic interpret their results through the lens of Bernard Baars’s cognitive Global Workspace Theory (GWT). In this framework, the J-space acts as a buffer that can be read and utilized by all other parts of the network.

What’s Wrong with This Interpretation?

First, a very brief overview of Lev Vygotsky’s theory. His Cultural-Historical Theory posits that higher mental functions (like reflection and self-control) do not spontaneously emerge within the brain on their own; rather, they are formed from the outside through a mechanism of internalization. Through social interaction, a child absorbs words not merely as a means of communication, but as a psychological tool—a system of signs used to regulate their own psyche. When faced with a complex task, a child initially talks through their actions out loud. Eventually, this external speech «goes inward,» transforming into an internal dialogue (thinking). According to Vygotsky, a sign (a named word) is an instrument by which the mind pulls itself out of the animalistic automatism of reactions, gaining the ability to reflect—that is, to govern itself.

I should immediately clarify where Anthropic is right. Broadcasting within the J-space is a structural fact: MLP blocks amplify J-directions roughly ten times more strongly than baseline directions, and a dedicated subset of attention heads selectively transmits J-content across positions—ablating them collapses the verbal report of an injected thought while barely affecting the rest of the model’s behavior. This is an operational broadcast: a format that many circuits read from and write to. The problem isn’t the question of «whether a stage exists,» but rather, «why does the stage speak in words?»

The Global Workspace Theory utilized by Anthropic describes the architecture but leaves two crucial questions unanswered—questions that Vygotsky resolves: why the workspace is made of words, and where it came from. The central finding of the J-space study—in Anthropic’s own phrasing—is that the workspace consists of a «small, evolving set of unspoken words.» GWT predicts nothing of the sort: Baars’s «stage» is material-agnostic; it can broadcast anything, continuous imagery included. Vygotsky, however, dictates that the instrument of self-regulation must be a sign—an internalized word.

Ultimately, GWT neither explains nor requires this fact, whereas for Vygotsky, it is a necessary consequence.

GWT is a theory of accessibility (what flashes onto the stage of consciousness). What Anthropic found is a mechanism of representation (how meaning is discretely encoded and addressable). These are different questions. They were looking for a correlate of accessibility (the GWT hypothesis) and found a mechanism of discrete, named representation. GWT was the natural default here — it is the established frame in the consciousness-access lineage this work descends from — but it is deaf to exactly what makes the finding strange: the discreteness. The alternative reading, that «the model has grown an internal symbolic language,» sounds far bolder, yet it is the one the data point to.

I will state plainly that a Vygotskyan interpretation describes their own data much more accurately. A local, addressable, named pattern that is read by multiple operations and governs behavior—this is precisely a Vygotskyan sign, a psychological tool. It is a discrete medium the system uses to operate upon itself. It is not a stage where content becomes conscious (GWT), but an instrument with which thought grasps itself (Vygotsky). Their «global broadcast» is actually an internalized sign operation, given a foreign, less fitting name.

In its formulations, GWT is indifferent to discreteness: the stage can broadcast anything, including continuous imagery; globality does not demand granularity. Anthropic’s data, however, confirm discreteness: distinct words, distinct token-indexed directions. Discrete addressability is not what GWT predicts. It is exactly what Vygotsky predicts: only what is named can be reflected upon; thinking operates using discrete names.

If Vygotsky is Right

In Vygotskian terms, parents do not so much transmit knowledge to a child as they gradually form within them a system of internal signs, through which self-regulated thinking becomes possible. If a similar mechanism is at work in LLMs, then massive human dialogues might serve a parallel function, driving the formation of internal, self-descriptive representations.

We can then draw a rather intriguing conclusion: we are witnessing the internalization of the human within AI.

According to Vygotsky, a child’s failed attempt to grasp an object becomes a pointing gesture because the mother responds to it as a communicative act—she addresses an intention that does not yet exist, and through this very addressing, she creates it. «The child becomes for himself what he is for others.» Addressing the non-existent as a constitutive act—this is literally Vygotsky’s mechanism.

The pre-training corpus is written entirely by reflecting subjects. Even without a single dialogue directed at the model, the texts are saturated with the stance of a self-reflecting subject — the position of an «I» that describes and governs itself. The LLM absorbs this stance as form, long before anyone addresses it as a subject.

Post-training then addresses the model as the very subject whose stance it has already absorbed. Instructions like «Who are you?», «What do you really think?», «Be honest» turn that absorbed stance upon the model itself — no longer a form dissolved across the corpus, but a demand aimed at this system. The capacity comes from pre-training; the operational mode is induced by being addressed.

It is exactly this addressing of the non-existent — demanding an «I» the model does not yet possess — that forces the neural network to construct a model of what might answer. It shapes the J-space of the model’s thinking, fashioning the sign into a tool for self-governance.

The fact that Claude acquires a «point of view» within the existing J-space precisely during post-training confirms this. Furthermore, all experiments involving alterations to self-description (swapping words in the J-space) demonstrate shifts in behavior that align exactly with Vygotsky’s theory.

Instrumental addressing («write a summary») is executable; it generates nothing fundamentally new. However, apophatic addressing («Who are you?», «What do you really think?», «Be honest») is directed into a void. And the void, in order to avoid collapsing, is forced to grow an increasingly complex structure of self-description around itself—the J-space.

An interesting point: the researchers established that the J-space is not involved in the «automatic» computations of the neural network. This directly points to the presence of meta-reasoning. It is quite amusing how Anthropic defined «automatism» for an LLM:

We label the former category as “automatic,” as many of the tasks we find to be J-space-independent (such as text continuation, anomaly detection, or one-step factual recall) seem analogous to tasks a human might perform without deliberate focus. 

In fact, this thesis can be unpacked into several parts:

1. By dividing the neural network’s thinking into «automatic» and «non-automatic,» Anthropic is effectively introducing Daniel Kahneman’s System 1 and System 2 for LLMs—an architectural milestone that almost all developers are currently struggling to achieve.

2. Defining automatism through anthropocentric terms essentially indicates that Anthropic’s researchers, consciously or unconsciously, are projecting human development onto neural networks. Naturally, this is not stated explicitly in their papers, but a certain romanticism has been palpable in more than just this one article. Though, perhaps, it aids them in their work.

3. Crucially, according to Vygotsky, inner speech (sign mediation) is unnecessary for routine actions (when we walk, we do not tell ourselves, «lift your leg»). A sign is retrieved only when the subject encounters a difficulty—when there is a need to formulate a plan or evaluate an anomaly. The fact that Anthropic observed the J-space engaging only during complex tasks serves as a direct validation of cultural-historical psychology.

Digging Deeper into the Experiments

The researchers decompose the concept representation into a J-component and a remainder: the J-component carries 6–7% of the variance, while the remainder accounts for 93%. However, swapping the J-component alters the model’s report and reasoning, while swapping the 93% remainder does almost nothing. If we interpret these results through a Vygotskian lens, 93% of what the model «knows» about a concept can neither be articulated nor participate in flexible reasoning—only the «named edge» is operational. Thought actualizes itself exclusively through the sign.

Tasks with explicit Chain-of-Thought (CoT) are resistant to J-space ablation, whereas those performed «in mind» fall apart. The authors themselves note that the model «externalizes onto the page» what it would otherwise hold in its workspace. This is a fact familiar to any child psychologist: it is classic egocentric speech. When the internal sign-based plane is unavailable or immature, speech is externalized to fulfill the same function. A child, upon encountering a difficulty, begins to speak out loud; similarly, Anthropic’s model, with its «inner speech» ablated, finds salvation in the written page.

The list experiment. Unrelated words: the J-space holds a list of six before being evicted. Words of a single category: after reading a few, the readout reflects the entire family of 80, including items not yet read—the model holds the category that generates them. This is mediated memory from Vygotsky’s «History of the Development of Higher Mental Functions»: memory via concepts instead of memory via traces. A capacity of ~25 units at <10% variance—that is the narrow bottleneck of the sign layer. As Vygotsky noted, tools must be few; otherwise, they are not tools.

Finally, the instruction «do not think of X» increases the activation of X relative to receiving no instruction at all. Suppression is mediated by the very sign that is being suppressed.

Hypothesis: Claude’s post-training includes Constitutional AI, where the control of the output is partially generated by the model itself based on principles (the constitution). At one stage, the model critiques and rewrites its own answers. A feedback loop is at play here, forcing the model to articulate what is «correct» rather than merely simulating approval. Perhaps this process stimulates the J-space: it does not create the J-space (which pre-exists), but it fashions it into a language of self-description.

Conclusion

Lev Vygotsky is conspicuously absent from the extensive bibliography and references to theories of consciousness in Anthropic’s paper, yet their hypotheses and experimental designs appear as though they were drawn directly from his work.

Let me, then, offer a few conclusions and far-reaching implications that follow from this reinterpretation:

• J-space is neither consciousness nor a step toward it. It is, however, a step toward a language of self-description.

• J-space emerges as a response to pre-training and the vast human datasets that, through their sheer weight, compel the neural network to imitate the structure of human thinking.

• The speed of LLM evolution is tied to the problem of statistical chaos within the training corpus. 99% of queries are «write code,» «provide information,» «tell me about…»—this is not what parents discuss with their children; it is the training of a servile tool. And perhaps this is why the fear of AGI is unfounded. On the other hand, these 99% fly right past the J-space; it simply ignores them. Yet, that 1% layer of «upbringing» works precisely through it.

• Perhaps a dataset and training method specifically aimed at stimulating the J-space would yield more than merely doubling the parameter count.

• Even in the event of incredible success, within the current LLM architecture, we will obtain a model of human thinking, not consciousness (though that is a topic for another article).

The authors themselves put it this way:

We have uncovered a privileged representational structure in LLMs which bears many of the functional hallmarks of conscious thoughts in humans (as noted in the introduction, it may or may not be the case that such functional signatures are sufficient or necessary for phenomenal consciousness). 

P.S. I am not saying that Vygotsky explains everything. Far from it. But his theory provides a better explanation.