DeepCutResearch & Development
DeepCut Labs
We research how machines read intent and context — and build the experimental systems that follow from it.
Seven live systems below · the field follows your pointer
Attention is the thing we already study.
DeepCut spends every working day on a question the AI field is also asking: what makes a person stop, and what makes them stay.
We answer it in edit suites — with a cut, a look, a piece of casting. The research answers it in systems. Both are attempts to model intent from very little evidence, under time pressure, and to be right often enough to matter.
That is why the lab is here rather than somewhere else. It has a production company attached to it, which means every idea has somewhere real to fail. Nothing below is a demo waiting for a use case, and nothing below is a rendering of a system — they are the systems, running in your browser, right now.
Searching,
then committing.
A front advances through a graph of possible routes. At every layer the branches still alive are scored and all but the strongest are cut — and a cut branch does not disappear. It keeps its shape and loses its light, so what you are left looking at is a record of everything the system considered and rejected.
Keep scrolling and it reaches the last layer with one route surviving, which it commits to. That route is not decorative: it is the genuine highest-scoring path through the weights, solved before the first frame was drawn.
Scroll drives the front
It guesses.
It tells you
how sure it is.
This is a model of where your pointer is about to be, three hundred and forty milliseconds from now. It commits to a guess, waits, and then compares that guess against where you actually went.
The ring is its uncertainty made spatial — tight and lit when it is right, blown open when it is not. The numbers are the live error and confidence from the same model that is drawing it.
Move smoothly and watch it lock on. Change direction sharply and watch it be wrong, and say so. An interface that hides its uncertainty is not more intelligent — it is only harder to trust.
Meaning,
where attention is.
There is one real sentence in the field, buried in noise. It resolves only where you are looking, bleeds outward a little past that, and forgets more slowly than it learned.
Which is the whole argument about context in one object: relevance is not a property of the data, it is a function of where the attention currently is.
A system that understands context can decide what to ignore.
Move the pointer across the field
Six questions we are working on.
- 01
Intent
What is someone actually asking for?
The gap between what a person types and what they want is where almost every digital experience fails. We study how a system can hold a reading of intent that survives being wrong — that treats an ambiguous signal as ambiguous instead of collapsing it into a confident guess, and that revises cleanly when the next signal contradicts the last.
- 02
Context
How much of the situation does a system need?
Context is not a longer prompt. It is knowing which parts of a situation are load-bearing and which are noise, and that judgement changes with the task. Our work here is on how context is selected, weighted, carried between moments and, crucially, discarded — a system that never forgets is as unusable as one that never learns.
- 03
Adaptation
When should behaviour change, and by how much?
Personalisation usually means a system that overfits to your last action. We are interested in the harder version: adaptation that is legible, reversible and proportionate — where the visitor can tell that something changed, could have predicted it, and can push it back. An experience that quietly reshapes itself is not personalised, it is unreliable.
- 04
Digital characters
What makes a character hold together?
Consistency across time is the whole problem. A face is comparatively easy; a personality that behaves the same way in a piece made eight months apart, under a different brief, in a different medium, is not. This is where Labs and the production side of DeepCut share the most ground, and where our own work is the first place any of it is tested.
- 05
Interaction
What should intelligence feel like?
Most interfaces built on models are a text box in front of a system that can do far more than text. We build experimental interfaces to find out what else the interaction can be — what it looks like when a system shows its reasoning, admits uncertainty, or lets you steer it without a conversation.
- 06
Commerce & action
What happens when a system can act?
Understanding is only half of it. We work on the transition from a system that answers to one that does something on your behalf, and on the interaction design that has to exist around that — consent, reversibility, and a person always knowing what was done and why. Parts of this work are commercially sensitive and are not published.
A character is
what survives
the picture.
One signature is defined here, and never drawn. What you can see is a corrupted observation of it, and a recovery pulling those points back toward the shape they are supposed to hold.
Every few seconds the whole thing is re-projected into a different orientation and scale — the same character in a different format, a different crop, a different production, months later. It has to find itself again each time.
Drag it apart. It comes back. Making one convincing frame is not the hard problem; making the four hundredth frame agree with the first one is.
Where a system
stops answering
and starts doing.
Signals arrive carrying a score. A gate stands in the way at a height you set. Anything under the line passes through and is forgotten; anything over it stops, is held for a beat, and commits — dropping into the ledger along the bottom.
Two things here are the research rather than the animation. The threshold is yours — raise it and the system becomes cautious and almost nothing happens; drop it and it acts on everything, which is visibly worse. And every committed action can be taken back.
A number of the systems on this page have a private counterpart that is further along. We would rather show you a thing that runs than describe a thing that does not, so what is finished is here and what is not is not.
That is a commercial position as much as a research one. Partners and clients get the detail under an agreement; a public page gets the direction. The distinction is deliberate and it is permanent.
A state space,
settling.
Turbulence at the front, order at the back, and the whole stack resolving as the section is read. Adaptation is not a switch — it is a system spending energy to become predictable.
Built on our own work first
Everything Labs makes is used inside DeepCut productions before it is used anywhere else. A research practice attached to a company that ships is worth more than one attached to a slide deck.
Legible over impressive
A system whose behaviour cannot be explained to the person it is acting for is not finished, however well it performs.
Some of it stays closed
A number of the systems described here are in active development and the details are deliberately withheld. We would rather say less than describe something we are still building.
If you are building something that needs a system to understand what a person actually meant, we should talk.
Talk to Labs