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

[Jeverettk]

This entire article rests in nonsense that fails to understand the basic mathematics. J-space is not discrete due to some ancient assumption about words and consciousness. J-space is discrete because the Jacobian lens used to infer the space can ONLY yield discrete answers.

There is no non-discrete solution, not because of Vygotsky, but because a J-lens cannot yield a non-discrete solution.

Similarly, there is no non-verbal outcome specifically because Anthropic focused the study only on verbalizeable representations. Again, not because of some poorly structured concept by Vygotsky, but because that is what was tested for.

I make no claims about Vygotsky here, only that the claim that Vygotsky's theory explains the J-space output is nonsense. Vygotsky is Vygotsky. J-space is discrete and verbal because that is the only thing the test could possibly have produced.

[Kamil_GR]

You're right about the search criterion, and the article says so: the J-lens is built over the vocabulary, so the alphabet is verbal and token-indexed by construction. Nobody disputes that. But that explains what the space is made of. It does not explain what it does — and that's the whole finding. Anthropic went looking for representations satisfying one property (verbalizability), and then found, in their own words, "rather surprisingly," that the same set satisfies four others they never searched for.

Three things break the artifact reading:

1. Ablating the J-space doesn't damage speech — it damages thought. With the J-space suppressed, the model stays fluent, parses text, classifies, and does one-step recall; multi-hop reasoning collapses. If this were merely "the channel of what the model is about to say," ablation would impair saying. It impairs reasoning instead, while leaving saying intact. A verbal probe's artifact cannot be selectively load-bearing for the non-verbal part of cognition. And note their own aside: explicit chain-of-thought survives ablation while the same problems solved "in the head" fall apart — the model, as they put it, externalizes onto the page what it would otherwise have to hold internally.

2. Selectivity rules out causality-by-construction. In their language experiment, the name of the passage's language appears in the lens at comparable rates across all four task conditions — but swapping it flips explicit report and flexible inference, while continuation and anomaly detection are unaffected. If the causal effect were an artifact of how the lens is defined, it would move everything downstream uniformly. It doesn't: the same representation is causally load-bearing for deliberate tasks and inert for automatic ones.

3. The load-bearing measurements don't use the lens. The categorical collapse at the workspace onset is measured by projecting the activation onto the line between the pure-A and pure-B activations — no J-lens involved. Early layers track the input mixture proportionally; from the onset layer on, the activation sits at one end or the other and flips sharply at threshold. Separately, their concept vectors and their two-hop probes are built by independent methods (mean activations over prompt sets), and the causal punch concentrates in the J-component even though the bulk of the variance lies outside it — and the remainder's residual effect is itself routed through the J-space, since clamping the J-coordinates removes it. The MLP gain on J-directions and the dedicated broadcast heads are facts about the model's weights, not about the readout tool.

On discreteness specifically: the lens outputs a softmax over the entire vocabulary — a distribution, not a one-hot. Nothing forces it to be peaked. And in the first third of the layers it isn't: by their own kurtosis measurement, readouts there are flat and uninterpretable. Peakedness appears at the workspace onset and fades in the final layers. A lens that "can only yield discrete answers" would be peaked everywhere and would have nothing to report about where in the network the structure lives.

So: the alphabet is discrete by construction. The bottleneck is not. That the network routes its deliberate reasoning through a sparse, capacity-limited code drawn from that alphabet, amplifies precisely those directions with its own MLPs far more than other directions, relays them with a specialized subset of attention heads, and carries the causal load on a small minority of the variance — while automatic processing bypasses the whole thing — was not built into the method. It was found.

And that conjunction is the Vygotskian claim. Not that Vygotsky explains the mathematics of the J-lens. That a discrete, named, capacity-limited layer which mediates deliberate reasoning, is bypassed by automatic processing, tightens under instruction, and whose loss is rescued by writing the steps out on the page — is the exact functional profile of internalized sign-mediated speech. GWT is indifferent to the medium: Baars's stage can broadcast anything, continuous imagery included. Vygotsky requires the medium to be a sign. That is a difference in predictive content, not in vocabulary.

(translated by Claude)