Causal AI
Positronic.
A physical computing substrate. The data is held in superposition, not in weights. Nothing is trained.
Eighty years of computing has moved one direction: from the transistor up through gates, machine code and languages to the token, each layer a symbol for the one beneath it. We went down a layer instead. Language describes three-dimensional objects, the relationships among them and the forces that act on them, so a sentence is built as the arrangement it describes and its field drives a frozen network of 362 coupled conductive loopsobeying only Faraday’s law and the RL circuit equation.
A body of text is not stored as a record. It is compiled into a key: the parameters of a field that returns the medium to the state the whole body produced. A question is driven into the same field at the same instant, and the answer is read from the interference.
The medium is a network of conductive loops coupled to one another through their own fields. Every loop both drives and receives. There is no input side and no output side, and the coupling among the loops is the computation. It has data, a great deal of it, but the data is not learned into weights. It is held across every loop at once, in superposition, and the physics settles it in the order the world settles things: cause before effect. That gives natural causal ordering and reasoning from first principles rather than from statistical association.
The same property opens a path to robotic proprioception. An actuator can sit on the medium directly, so the body it moves writes its own state back into the field, with no controller between the computation and the motion.
Alongside it, a line of work showing that relations behave as group elements: composition as arithmetic rather than inference, extrapolating far past training where a transformer three times the size falls to chance. Provisional patent applications are on file.
Record
- 2026 · August
First preprint: language computed by the cycle topology of neural tissue, with no learned parameter.
- 2026 · September
Superpositional memory demonstrated on a 362-loop artificial substrate. Provisional patent applications filed.
- 2026 · October
Fourth preprint: the lattice’s governing equation derived from Maxwell’s equations and checked against the 362-cavity substrate.