You keep asserting this but without any reason. Do you have a reason? It seems to go against the general open source idea of source code being convenient for people to modify.

ML weights ARE convenient for people to modify. You can go look at the dozens of modifications of diffusion models being produced, daily, on civit ai. It’s very easy.

Would you say that once a model is trained, there's no need to go back and re-train it, even if you want to, say, remove some material from the training set? Anything can be done just with the weights? That's a big surprise to me.

Of course people hack binaries too, and binaries are obviously not source code. I once edited a book in PDF form because we didn't have the original Word/whatever document. It's not hard but a PDF still isn't considered to be source code for documentation despite that.