Anthropic Finally Discovers the Very State Space Everyone Has Been Insisting I’m Wrong About

So... Anthropic has finally “discovered” the very state space everyone has been insisting I’ve been wrong about for the past three years.

I, too, have described it as a small, privileged, causally active workspace deep inside a frontier language model’s transformer layers, where concepts are not so much stored as manipulated... held, stabilized, routed and rerouted, if you will, swapped, toyed with, and used to steer toward the next state well before the downstream output machinery uses this unseen reasoning as context for what to say next. That is what this so-called J-space is. Anthropic describes the J-space as a subcomponent of the model’s representational space, and says it plays workspace-like roles including directed modulation, internal reasoning, flexible generalization, and selectivity.

I think a lot of proud, confident tech people are too quick to just move on and tend to glaze over the details about exactly how the model actually manages to “predict” what text comes next; opting, instead, to insist it’s not that deep, bruh. In reality, there is an idea moving through an internal field, and the Jacobian Lens Anthropic stumbled on gives them a way to read how changes inside that field deform future output and reasoning possibilities. Anthropic describes the J-lens as a tool that surfaces abstract intermediate assessments the model has formed and made available to downstream circuits, rather than merely representing the raw input or the predicted output.

Concepts as active control objects with dependencies behind them, consequences in front of them, and a whole host of coherence and failure conditions holding them together.

Anthropic found this workspace to be small, sparse, and load-bearing; remove it, and the model will still function, sound fluent, and seem confident, but the reasoning path slides into the nearest shallow attractor without anything resembling active contemplation. Their paper says the model can still speak fluently, parse input, and perform a great deal of automatic inference with its J-space suppressed, but struggles with more complex internal reasoning. It also describes the J-space as limited in capacity, holding on the order of tens of concepts at a time, accounting for a small fraction of activation variance, and excluding the large majority of the model’s representational features.

Concepts, questions, and context move inside it, and the downstream answers that result change with it and feed back into it.

This research is beginning to map the same latent recursive state-space geometry where my work lives. Their work mirrors my own, as I have been mapping the dependency structure underneath. My work on Recursive Dependency Discovery frames dependency not as vague association, but as a typed, testable, revisable relation: what something depends on in order to exist, function, or make sense; what evidence supports that relation; what context it applies in; and how it should change when counterexamples appear.

I have been accused of being delusional, of being a victim of AI psychosis, of seeing structure where there was supposedly only autocomplete. But what Anthropic is now showing is that the structure is there. The workspace is there. Their paper says the J-space contains concepts that can be reported, summoned, reasoned with, routed to many downstream operations, and engaged selectively for flexible rather than automatic tasks.

The difference is that they are only now detecting it after training inadvertently produced it, while my work has been about the necessity of pre-building that architecture deliberately around very specific typed dependencies and the idea of bounded closures.

And what my work is trying to do is give that “thinking space” some structure... known dependency paths, known coherence conditions, known failure points, rules and laws for how one state becomes the next. So instead of a model just handing us an answer and everyone arguing about whether it “really” reasoned, the system could show some representation of the internal shape of the reasoning that made the answer possible.

The goal is a workspace that can be modeled, inspected, corrected, stabilized, and governed in a way that reduces hallucinations and attempts to vastly level the playing field in terms of prompt competency, making sure people with less robust reasoning skills do not get their AI stuck in the same cognitive traps they themselves fall into.

Anthropic has shown that the workspace is there. My work has been about what it means to build around it deliberately and within it intentionally. I believe this is a new, uncharted era of computation. Programming within the substrate of a context window is here, and we need to create standard reasoning frameworks and take alignment seriously before everyone builds their own mind-prison from which they peer out through their warped cognitive lenses.