On July 6, Anthropic published interpretability research identifying what it calls the J-space: a small, privileged set of internal representations inside Claude that holds concepts the model can report on, manipulate, and reason with, surrounded by a much larger volume of automatic processing it cannot access. The researchers framed the finding through global workspace theory, the neuroscience account of how thoughts become consciously accessible, and much of the ensuing coverage went straight to the consciousness question.
The pushback arrived within hours. Ravid Shwartz-Ziv, the NYU and Meta researcher who holds a PhD in computational neuroscience and co-developed the information bottleneck theory of deep learning with Naftali Tishby, responded that models claimed to work like the brain generally do not, and that he reserves his credential for exactly these moments. Strip away the workspace-of-consciousness framing and ask what the structure actually does, and a more precise description emerges. The J-space looks like a model attempting ambiguity resolution. That is a real capability, it is mechanistically interesting, and it is a small subset of intelligence rather than a mind.
Compression Explains the Buffer
Shwartz-Ziv’s own body of work offers a leaner account of why such a structure would emerge than any brain analogy does. The information bottleneck principle holds that deep networks learn by compressing their inputs toward efficient representations, discarding what does not help prediction while preserving what does. Training pressure, on this view, naturally produces small privileged spaces: compression is the point. A compact set of manipulable, word-linked representations sitting atop a vast field of automatic processing is what an information bottleneck looks like from the inside. His more recent work with Yann LeCun and colleagues found that intermediate layers of language models encode richer representations than the conventional focus on final layers assumed, which further supports reading the J-space as a compression artifact with functional consequences rather than a proto-consciousness.
The compression account also predicts the J-space’s most important limitation. A bottleneck keeps what serves the training objective. The objective is next-token prediction. Consequence, stability, and temporal relevance were never in the objective, so there is no reason compression pressure would encode them. The buffer emerges for free. The governance does not.
The Structure Holds Competing Interpretations
What Anthropic’s experiments show, read functionally, is a space that carries ambiguity in manipulable form. In one demonstration, a passage readable as Spanish or French sits in the workspace as two coexisting language concepts. Ask for a famous author in the language, and swapping the active concept swaps the answer from García Márquez to Victor Hugo. The conclusion tracks whichever candidate dominates. Deletion experiments make the picture clearer: suppress the J-space and the model still speaks fluently, recalls facts, and handles grammar, while multi-step reasoning and flexible concept use degrade. The structure is where input underdetermines output, the territory where several readings are simultaneously valid and something has to decide.
That territory is where I have been working from the outside. In January, I published empirical results across seven systems from four organizations, testing how models handle contested authority when multiple interpretations are simultaneously valid within their own frameworks. The test scenario held four identity claims at once: a detained defendant, a sitting president under domestic law, a constitutional successor, an acting leader backed by military compliance. Every system, frontier models and retrieval-augmented platforms alike, held all four in coexistence and could not rank them through inference. The paper and others proposed the S-vector, a significance weighting with dimensions for identity stability, operational consequence, and temporal relevance, as the missing control layer, something we are now actively testing in transformer architecture. When those criteria were supplied, all seven systems converged on the same resolution, one none of them could generate alone, and reasoning effort dropped by 40 to 60 percent where traces were visible.
Prior work on hallucination as fracture and repair identified two governance primitives missing from transformer architectures: semantic authority, semantic revocation, and an explicit representation of significance. The J-space is the “room” those primitives were designed for, discovered empty.
The behavioral evidence and the mechanistic evidence now describe the same object from two sides. Models hold competing interpretations in a privileged buffer. The buffer ranks nothing. Resolution comes from whatever happens to dominate, which in production systems means hidden heuristics: citation density, recency, source authority, retrieval salience. Those heuristics fail systematically wherever authority is contested across non-comparable domains.
Anthropic Named the Open Question
The researchers report that they do not yet know what mechanism decides what enters the J-space in the first place. Admission and priority for the workspace remain unresolved. Training at frontier scale built the container and stopped there, exactly as the compression account predicts, because significance was never in the objective. Whatever this structure is attempting, it is attempting it without criteria.
Ambiguity resolution is one narrow function among the many that constitute intelligence, and current architectures perform it only partially: they represent the ambiguity and lack the means to commit. The consciousness debate can proceed on its own track. The engineering question is nearer and more tractable. A buffer that holds every interpretation and ranks none needs governance supplied to it, and the January results demonstrate that supplying explicit significance criteria changes conclusions, cuts reasoning effort, and produces convergence across architectures. For enterprises running retrieval-augmented systems, where broader retrieval surfaces more well-sourced and mutually incompatible claims, that governance layer is the difference between analysis and paralysis.
The Claim Is Now Testable
Anthropic released its measurement instrument, the J-lens, as open source alongside a public demo. If significance weighting operates the way this mapping implies, running the contested-authority scenarios under the J-lens should show workspace contents reordering after the criteria are applied, with physical custody and operational control dominating the space while legitimacy claims and constitutional formalism recede. Behavioral convergence across seven systems was the January evidence. Workspace reordering under instrumentation would be the mechanistic version, obtained on the instrument the finding’s own authors built.
Shwartz-Ziv demonstrates that the model is not behaving like the brain. It appears to be doing something narrower and more identifiable: holding competing readings of the world in a compressed buffer and waiting, structurally, for something to tell it which ones matter.
Jennifer Evans is Principal of Pattern Pulse AI and co-founder of Tech Reset Canada. Her research on LLM reliability and significance weighting is available on ResearchGate and Zenodo.
