Piling on/agreeing with you: The “magic smoke” isn’t the model, it’s the infrastructure the model plugs into, the data->feature pipeline, AND the model. Assumin

Piling on/agreeing with you: The “magic smoke” isn’t the model, it’s the infrastructure the model plugs into, the data->feature pipeline, AND the model. Assuming you’be done the things you mention (and maybe with one additional assumption of several models operating at once), I would also consider the models themselves to USUALLY not be super super sensitive.