Would the use of a coding AI be permissible if it were only trained on language syntax, library information, and possible (optional) FOSS source code?

You would end up with this if you had no sample code: https://en.wikipedia.org/wiki/Duff's_device

Anyway, permissible for what?

Codeberg just went fully AI adverse. I am a retired guy and I don't want to spend the time to code every jot and tittle of a big project - I wrote my own basic interpreter (I specified, claude wrote) If I did not have the AI help I would be years doing this - I may not have years.

So by permissible I am talking about an AI that can generate code to spec that is free of copywrite objections.

Why are you asking here and not Codeberg then? Nobody can answer for them.

There's no way you're going to be able to train an LLM from scratch without a giant corpus of real code and English text. Even if you managed to procure that ethically, it would only address one of a number of objections people have to these things.

This is why I am asking here - the objections are not only with Codeberg - I will probabley maintain my own git instance (once I learn how) but I am asking programmers at large. Codeberg has it's policy - and that policy was driven by some of the programmers there. so I am asking programmers, not codeberg.