How it works
The geometry AI thinks in.
A manifold is a learned map of a domain, with a position on it. Pair it with fast System 1-type models and a language model, and you get AI that reasons over structure instead of guessing from a prompt.
Three layers, three speeds
Each layer does the job it is good at.
The manifold
slow · learned · persistent
Where are we, and where should we go next?
System 1
fast · calibrated · typed
What happens now? On every turn, in milliseconds.
The language model
fluent · flexible · steered
How do we say it?
The manifold
- Answers
- Where are we? Where next?
- Speed
- Changes slowly
- Output
- Coordinates, uncertainty, a path
- Remembers
- Everything that matters, across sessions
System 1
- Answers
- What now?
- Speed
- Milliseconds, every turn
- Output
- A typed judgement: a probability, a label, a choice
- Remembers
- Nothing; reads the manifold each time
The language model
- Answers
- How do we say it?
- Speed
- Seconds
- Output
- Words, voice, explanation
- Remembers
- Only what it’s given
Why not a language model alone? It has no persistent sense of where someone is. It makes the same judgement differently on different days. And it is too slow and costly to consult on every small decision.
Why not a manifold alone? It knows, but it can't talk.
The combination gives the language model something true to speak from, and a fast layer that keeps its decisions consistent.
The geometry
Every state is a place.
Every state worth telling apart is a point. The distance between two points is the cost of moving from one to the other; the metric tensor encodes it. Where movement is hard, the space curves. The best route is a geodesic: not the straight line, but the path of least action through the curvature.
See the geometry move: the learning manifold in eight screens →
A point is…
In learning
a learning outcome
In emotion & personality
an inner state
Curvature is…
In learning
difficulty: dense prerequisites, big leaps in abstraction
In emotion & personality
resistance: how hard it is to move from wariness to trust
A geodesic is…
In learning
the learning path that picks up nearby mastery on the way
In emotion & personality
how a relationship can develop without lurching
A position is…
In learning
where this learner is, dimension by dimension
In emotion & personality
how this character feels, and who it is
Operators
Four questions the geometry can answer.
Each operator earns its place by answering a concrete question that a product needs answered.
Entropy
“How uncertain are we about this position?”
Look where uncertainty is highest: the next question to ask, or the next thing to attend to.
Divergence
“Is the trajectory drifting from where it should go?”
Correct course early, before drift becomes distance.
Jacobian
“If this moves, what else moves with it?”
Prioritise high-leverage changes. Mastering fractions moves decimals, percentages and ratio with it.
Green’s function
“How far does one event ripple?”
Size the response to the event, neither too small nor too large.
Bayesian framing
A complete Bayesian system.
Prior
The manifold
Structure distilled from authoritative sources. For learning, that means curricula, textbooks and international frameworks. For character, it means established models of personality and affect. Computed in advance and tuneable to a culture.
Likelihood
Evidence
Each observation updates belief about a hidden state. In learning, that is an answer, a spoken sentence or a marked script. In character, it is a turn of conversation. This is standard, well-understood sequential inference, and the operators act here.
Posterior
The trajectory
Where this person or character is, where they’re heading, and what should happen next. It is what the product renders.
The framing isn't cosmetic. It's why the operators are well defined: they act on a posterior over a manifold, not on a list of topics. It's why a new curriculum, or a new character, doesn't mean starting again: each is a chart on a shared prior. And it's what we are. Knowledge graphs are discrete and brittle; a manifold is continuous and learned. Retrieval has no model of the person. Language models are powerful components, but they aren't the asset. The inference is textbook. The trajectory belongs to the person. The manifold is ours.
What you get
The geometry, without having to hold it.
The mathematics is public. The calibrated geometry (the specific metric, weightings and orderings) is what we build and keep improving. Licensees run the manifold locally, on their own infrastructure: the calibration stays ours, the data stays theirs. Priors set the geometry; evidence refines it locally, inside each customer's own deployment. We never touch the data.