For a decade, progress in AI has followed a simple rule: more data, more parameters, more compute. It has worked remarkably well. But every step up the curve costs more energy, larger data centres and longer training runs – and once trained, the model is frozen.
Physics Network Intelligence turns the question around. Instead of asking how large a model must be, PNI asks how much structure a given energy budget can sustain. Memory is a graph under a fixed budget; what no longer pays off is forgotten by rule. Learning does not stop after training – it continues in operation, from every signal.
This makes the energy budget a control you can turn at runtime: the same principle runs on a robot chip or a server, and keeps working at its optimum when resources change. We believe this is where the decisive questions of the next years lie – and it is what we are building.