Teaching machines the shape of things.
Every model is a portrait of what it was trained on. The large language models of the last few years were trained on nearly everything humanity ever typed, and they became fluent in it: our arguments, our recipes, our code, our jokes. Then, more or less, they ran out of internet.
Compute keeps compounding. Architectures keep improving. But a model can only learn what it is shown, and the easy data is gone. The next leap in artificial intelligence won’t come from scale alone. It will come from new kinds of data, and from new places for models to learn by doing.
Intusent Labs builds both.
The unwritten world
Language models learned to write because we wrote everything down. Billions of people, over decades, poured their thinking into text and left it where a crawler could find it.
The world of things was never written down like that. How a bracket carries its load. Why a hinge is chamfered here and not there. What an engineer sees at a glance that no drawing ever says out loud. That knowledge lives in files that never leave the companies that made them, on factory floors, and in the hands of people who build. It is one of the largest bodies of human expertise that today’s models have barely seen.
We call it the unwritten world. It is where everything you can touch was decided. If AI is going to change the physical world, this is where it has to learn.
Geometry is easy. Intent is not.
A file can describe the shape of a part to a thousandth of a millimeter. It almost never says why. Why this wall is 1.2 mm thick and not 1.0. Why the holes sit exactly where they do. What the part has to survive once it leaves the screen.
- sketchbase profile · 80 × 40 · fully constrained
- extrude12.00 mm
- shellwall 1.20 mm
- hole ×4⌀ 6.6 through · M6 clearance
- filletr 2.00 · 8 edges
- whynot recorded
That why is the difference between a picture of a machine and an understanding of one. It is also where today’s models fall apart: they can describe a gear in perfect prose and still not know what it is for. So that is where we spend our time: capturing intent at the source, with the people who hold it, and turning it into signal a model can learn from.
Apprentices, not librarians
Reading about a bridge has never taught anyone to build one. Pretraining is reading. What comes next is practice.
The most capable models of the next decade will learn the way apprentices do: attempt, get checked, get corrected, try again, millions of times over. That takes environments with real rules and real consequences, where a design either fits or it doesn’t, and where a reward actually means something. We build those environments. Patient ones, exacting ones, honest about failure.
Starting with the shape of things
We are beginning with geometry: the quiet grammar every manufactured object is written in. Sketches and constraints, features and assemblies. The phone in your hand, the chair beneath you and the engine on a wing were each, at some point, a file and a long series of decisions.
It is a demanding place to start. Geometry is unforgiving; there is no roughly right about whether two parts fit. That is exactly why we chose it. A model that can reason about tolerances, loads and fits can learn to reason almost anywhere matter is involved, and every answer it gives can be checked.
What changes
Picture the day a model understands objects the way it already understands sentences.
You describe what you need and get back a part that fits, carries the load and can actually be made. An agent explores ten thousand variations before you finish your coffee. A five-person workshop designs like a five-thousand-person company. Repairs are designed for the broken thing in front of you, not the nearest catalogue entry. Robots finally understand the objects in their hands. Things get lighter, stronger, cheaper and less wasteful.
None of this replaces the people who make things. It makes their judgment the most valuable training data in the world, which is exactly why we work alongside them.
The first wave of AI learned from everything we wrote. The next will learn from everything we make.
Inside the lab
Intusent Labs is small, deliberate and a little obsessive. We think what a model learns from deserves as much care as the model itself, so we hold ourselves to a few things.
- Signal over volume. Ten thousand examples that are right beat ten million that are almost.
- Provenance, always. Everything we make can be traced to where it came from and who shaped it.
- Measured, not claimed. If we can’t measure what it teaches a model, we don’t ship it.
- People, not raw material. The experts whose judgment we capture are collaborators, and are treated like it.
- Hard problems first. We go where close enough isn’t.
Write to us
If you’re training models that need to understand the physical world, or you’ve spent a career making things and think AI should learn from people like you, we’d like to hear from you.