J-Space and the Geometry of What Matters

J-Space may be the workspace through which an intelligent system makes selected information globally available; AgileBrain may provide the latent motivational geometry through which that system represents what matters.

Our Chief Scientist, Dr. Bill Nolen, has written a paper proposing that the same twelve-need structure AgileBrain^® measures in people may also describe how large language models represent what matters in artificial intelligence. Here's why I think it could be one of the most important ideas we've put forward — and how anyone can help test it.

Motivation is the Missing Model in AI?

Every intelligent system, human or artificial, faces the same basic problem: it is flooded with far more information than it can act on, and it has to decide what matters. Not what is true, or what is nearby, but what is significant — what should be pursued, protected, repaired, resisted, or avoided. Psychology has a name for the machinery that answers that question: motivation. AgileBrain is our attempt to systematically map human motivation (Pincus, 2022). And it's showing great power in use among students, veterans, caregivers and the general population.

We built that map from human data. What Bill's paper asks is whether the same map might describe something we never designed it for: the inside of an AI.

A workspace inside the machine

The paper builds on recent interpretability research (Gurnee et al., 2026) that identified something remarkable inside large language models — a small, privileged set of internal representations the authors call J-Space. Only a sliver of everything a model computes lands there, but what does has a special status: it can be reported, deliberately manipulated, used as a step in reasoning, and made available to many downstream processes at once. In other words, it behaves like the "global workspace" cognitive scientists have long used to describe the difference between the vast parallel processing going on beneath awareness and the narrow, shared stage where deliberate thought happens.

Finding a workspace, though, raises a harder question — one the original authors explicitly left open. What decides which representations earn a place on that stage, and how is the content organized once it's there?

A motivational geometry

Bill's proposal is precise, and deliberately narrow. He is not claiming that everything in J-Space is motivational; most of it — locations, numbers, grammar, facts — plainly isn't. The claim is about the subset that carries agent-relative significance: the moments when the system is representing not just what is happening, but what it means for someone. When a model reasons about a locked door, an overbearing boss, and an arbitrary government rule, those situations share almost no surface content — yet all three turn on the same motivational relation: restricted autonomy. The hypothesis is that this shared meaning isn't just a convenient label we impose from outside, but actually a real, recoverable structure in the model's internal geometry — structure that lines up with AgileBrain's twelve dimensions, their promotion and prevention directions, and their relative intensity, which we call emotional energy.

Strikingly, the interpretability findings that already exist strongly point this way. Models represent safety when a described drug dose is normal and danger when it's an overdose, before generating a word. Concepts of grief and sympathy surface when a user mentions a loss in passing, ahead of any reply. Conflict, ethics, and preference all appear as workspace content that causally shapes behavior. These aren't proof, and Bill is careful to call them motivating observations rather than evidence. But they're exactly the kind of thing you'd expect to see if the hypothesis were true.

A theory that can be proven wrong

What I find most compelling is that the whole proposal is falsifiable, and Bill lays out precisely how to break it. The strongest test is what he calls blind recovery: strip every motivational label out of the stimuli, orthogonalize the semantic content so topic can't stand in for motivation, read the structure straight out of the model's internal geometry, and ask whether the four domains, twelve needs, and promotion/prevention directions reappear on their own — measured against serious competitors like valence-arousal, basic emotions, regulatory focus, and established appraisal theory. If they don't, the theory loses. That's the point. A framework that can be embarrassed by data is a scientific one.

Thinking bigger

Reading it, I couldn't stop the idea from running further than the paper deliberately does — so let me put my own speculation on the table, clearly labeled as speculation.

If a system's motivational representations are what tell it what matters, then motivation may be more than a passenger in cognition — it may be part of what makes general intelligence go. A workspace gives you a stage; motivation gives you a reason to put something on it. It supplies intent and urgency, directing attention and resources where they're most needed. That is close to the role emotion plays in Antonio Damasio's account of mind, where the felt regulation of the body's condition — homeostasis — is what binds an organism's scattered processes into a single, caring point of view. Motivational force may be the ingredient that turns a powerful predictor into something that genuinely attends. I don't say that's true. I say it's now a question worth asking out loud.

If motivation does play that organizing role, it should come with a natural order of priority — a triage, the way one's emotional needs quiet or clamor depending on what's at stake. My guess at the ordering runs like this: foundational prevention first (keep the floor from falling out), then experiential prevention together with foundational promotion, then aspirational prevention with experiential promotion, and finally aspirational promotion. Threats before gains at every level; the ground floor before the top. It's a diagonal sweep through the level-by-direction AgileBrain grid — and it makes a prediction you could actually check against workspace priority.

There may be a second axis of urgency, running along what I think of as decentering — how far a concern reaches beyond the self. The four foundational needs, one per domain, line up almost too neatly: Safety (the self), Autonomy (the self acting on the world), Inclusion (the self among others), Justice (the world as it should be, whether or not it touches me). Concern seems to travel outward along that path, and to grow more abstract as it goes.

That connects to the paper's fifth prediction — cross-agent generalization, the idea that the same motivational structure should appear whether the concern belongs to the user, a third person, or an organization. To me that's really a question about the unit of analysis, and it behaves like a set of concentric circles of support. Whose need is at stake — the individual's, the family's, the team's, the neighborhood's, the faith community's, the region's? The further out the circle, the more abstract and the more "spiritual," in AgileBrain's sense, the motivation becomes, shading from lived social bonds toward impersonal ideals. Where a concern sits in those circles almost certainly shapes how it's prioritized.

An open invitation

Here's what excites me most. Because an artificial workspace can be read, prodded, and causally intervened on in ways a human brain can't, these are questions we can actually test — not eventually, right now! And they don't all have to be tested by us. The most convincing version of this evidence would be crowdsourced: many hands, many models, many scenario sets, probing the same predictions in the wild and reporting back. If the motivational geometry is really there, it should turn up again and again — in labs that have never heard of us.

So consider this an open invitation. If you work in interpretability, in psychometrics, or in AI evaluation — or you're simply the kind of person who enjoys trying to break a good hypothesis — Bill has handed you a clean target and a clear way to aim at it. We'd love to know what you find, whichever way it cuts.

— JD Pincus, PhD

References

1. Gurnee, W., et al. (2026). Verbalizable Representations Form a Global Workspace in Language Models. arXiv:2607.15495. arxiv.org/abs/2607.15495
2. Pincus, J. D. (2022). Theoretical and Empirical Foundations for a Unified Pyramid of Human Motivation. Integrative Psychological and Behavioral Science. Read the study · doi:10.1007/s12124-022-09700-9