Neuralzoom is a closely-held technology house working at the edge of artificial intelligence, applied algorithms, and robotics. What we build, where we deploy it, and who it is for — remain, by choice, our own affair.
We are a small house, intentionally. We build instruments — models, algorithms, and machines — that attempt to make sense of the world where the world is least willing to be understood. What we discover, we keep. What we keep, we compound. What compounds, we return, in quiet and deliberate ways, to work and people we believe in.
Our work lives in three rooms. They share walls, ideas, and on most days the same coffee. The details of what each one is working on are — with rare exception — held in-house.
Long-horizon work on reasoning under uncertainty — how machines form beliefs, revise them, and carry the right amount of doubt when the world refuses to cooperate. Our papers stay on our shelves.
We write the kind of algorithms a patient mathematician writes once and refuses to publish. They run against some of the hardest problems we know, at a cadence the outside world never sees.
Embodied systems that sense, decide, and act. We are interested in machines that behave well when no one is watching them — and behave better still when everyone is.
Probabilistic forecasting systems, built to reason in the grey. They estimate what is likely, what is possible, and — most usefully — what they do not know. What they are pointed at is not for this page.
A self-owned fabric of silicon, storage, and software. Our models train, our robots learn, and our engines run inside a perimeter we draw ourselves. Nothing of ours needs to leave the room.
Tooling, interfaces, observability — the dull, essential surface between human judgement and machine output. We spend more time here than we expected to, and consider it time well spent.
A rotating list of the problems our house is sitting with. Titles only — specifics, methods, and findings remain internal. Consider this a window, not a door.
What happens to a hard-won regularity the moment it is noticed by the thing it was found in? A question we are still sitting with.
Every serious system behaves, eventually, like the worst minute it has ever had. How to write code that loses that argument gracefully.
On designing embodied systems whose incentives remain aligned even in the absence of an observer. Notes from a long, quiet corridor.
We are interested in machines that know the size of their own ignorance. This note sketches what we think that costs, and why it is worth it.
Most of what calls itself an interface is a dashboard. We try to build instruments instead — things a careful person can rest their judgement on.
A short essay on the capacity curve of serious work, and the point at which adding people, capital, or attention begins to destroy it.
Houses develop cultures whether they design them or not. We chose to design ours, up front, and in writing.
Serious work unfolds on horizons that reward the patient. We measure ourselves in years, not quarters, and act accordingly.
We do not hire for growth. We hire, if at all, for depth. A house of few careful people is our ceiling and our preference.
We do not publish strategies, do not court attention, and do not comment on work that has not yet been done. Quiet houses stay curious.
Our systems produce probabilities; our people carry responsibility. We do not delegate judgement to things that cannot be held accountable.
Every thesis, every mistake, every retired idea is archived. Institutional memory is the one moat a small house can actually afford.
What exceeds what the house needs to operate is not ours to keep. A share of each cycle is committed, in advance, to public good.
We treat responsibility as a working line item, not a page in a brochure. A portion of each cycle is committed, in advance, to three areas we can measure and stand behind.