I'm far from an expert on AI or anything like that, but here's my understanding: The "regular" ECMWF model is based upon dynamic and thermodynamic equations (speaking very simply here), with a set of initial conditions and observations to start from. Those equations are evaluated through time to obtain forecasts. The AI version is based upon some collection of past "training data" (such as previous atmospheric conditions and observations, and perhaps analogues), which are sort of "regressed" or "correlated" against an independent set of observed data (think of it as a highly sophisticated multi-linear regression, in a way). Others can chime in, I hope this is reasonably accurate of a description albeit kind of simple.