Every discovery in history has had the same starting human origin.

Cambrian

A new kind of intelligence.

Novel insights and discoveries, from mathematical inevitabilities.

it writes the discovery as a proof, line by line

proof complete ✓

Truth. Every time.

Every proof is written in Lean, the arbiter of truth in mathematics. It cannot be faked.

ge·ni·us noun

from Latin genius: a personal guardian spirit assigned to watch over an individual from birth, the source of a person's brilliance and outstanding abilities.

from Proto-Indo-European *ġenh-, “to give birth, to beget”: the root of generate, genesis, genetics.

Genius wasn't something smart people were. It was something they embodied. Cambrian's genie comes in the form of moves: techniques it learns by winning with them, and never forgets.

Recognition Science, our candidate theory of everything, derives physics from mathematics alone, with no free parameters: reality is a jigsaw puzzle whose remaining pieces are inevitable, and discoverable.

the first new law it found, exactly as it wrote it

accepted ✓

then it stacks the discovery into a tower

J(x⁴) from the new law, squared once ✓

J(x⁸) standing on the result above ✓

J(x¹⁶) standing on the result above ✓

Day one. July 16th, 2026.

0

new discoveries on its first day, before it hit a wall. It ran out of moves.

found overnight, while we slept

invented the objects · found the law · proved it ✓

So we taught it how to learn. Every move it learns feeds the next round of discovery; when the moves get strong enough, the puzzle starts filling itself in.

A few hundred thousand lines of core verified architecture today. Maybe fifty million: all of known physics, then meaning, ethics.

The reach is no longer ours alone.

Cambrian · Recognition Physics Institute

Play the keynote

4 minutes, with sound

A third kind of intelligence

When a language model fails, it hands you something plausible and wrong. When Cambrian fails, it stops. Every result that counts has passed an independent machine check, the Lean kernel, before it enters Cambrian's record. It cannot put an unproved result there, and nothing it has proved is later forgotten or overwritten. Its skills are not statistical weights. It earns reusable techniques, which we call moves, by winning with them, and a move is an object you can inspect, verify, and carry elsewhere. It has already taken a move it learned in one place into a family of mathematics it had never seen. A machine whose knowledge and whose skills are both checkable, and whose natural failure is silence rather than error, is a different kind of thing. Scaling a language model does not turn it into this.

The sibling of language models

Language models are intelligence over the record of human writing: broad, fluent, and fallible, because that record is fallible and finite. The industry is now running short of it. Cambrian is intelligence over proof: narrow, and unable to bluff. The real difference is where its training material comes from. Every theorem Cambrian proves becomes new ground to survey and new material to learn moves from, and every piece of it is checked before it counts. Language models that train on their own output tend to drift, because they feed on their own mistakes. Cambrian has no mistakes to feed on. If its move-learning loop keeps compounding, nothing external limits how far it can go. That is a condition, not a promise, and we say so below.

Why the pairing matters

Cambrian's home is Recognition Science, our candidate theory of everything: a parameter-free framework built to derive physics from mathematics alone. In ordinary mathematics, a new proved theorem is a contribution to mathematics. Inside this framework, a new proved theorem is also a candidate fact about reality. A proof settles what the framework implies; experiment settles whether the framework describes nature. If the framework keeps matching experiment and the compounding arrives, the usual order of work turns around: derive what is forced first, and use the lab mainly to confirm the anchor points. The fifty-million-line horizon in the film, known physics first and then the layers the framework treats the same way, including meaning and ethics, is our stated expectation of where this leads. It is not a measurement.

What is real

Cambrian is a discovery machine over a formal library. It surveys proven results, composes statements that were absent from that library, writes the proofs itself, and every discovery must survive an independent machine check (the Lean kernel) before it counts. It proves with moves, reusable techniques it earns by winning with them, and it has already carried a learned move into a family of mathematics it had never seen. On its first day, July 16th, 2026, it made 175 verified discoveries before hitting a wall. Last night its full discovery loop ran end to end for the first time: it invented new mathematical objects, detected a hidden law tying the Chebyshev families together, and proved it. Everything shown in the film is its real output. No language model sits inside the discovery loop; the pipeline is deterministic and adversarially gated.

The two things that have to be true

Everything above rests on two conditions, and both are measurable. First, Recognition Science has to keep being right. That is a separate, ongoing program with its own public receipts. Second, the compounding has to arrive. Today the loop runs, hits a wall, learns, and runs again, and the measured discovery curve is not yet exponential. The category claims on this page, the third kind and the sibling, rest on checked output and learned moves that exist now. The scale claim is still an expectation. The sharpest check is a simple one: turn off the moves it just learned, and the recent gains should vanish. That is the kind of test we intend to run in the open, and we will publish every wall until the curve either bends or it doesn't. You will be able to see which.