That and re-training the models once the code is rewritten. Presumably MS doesn’t have access to the same training sets (they almost certainly have their own, but we don’t know the breadth or quality by comparison to OpenAIs data sets)