There are significant differences. At a very high level, AI models are trained on reanalysis data. It is analog/pattern recognition on steroids. I think there are some traditional modeling-like guard rails involved, but others that have more knowledge should chime in. Traditional models are physics-based, solving numerical fluid dynamics equations and then adding in parameterizations for things that can't be realistically calculated (and some tuning for good measure).