Am I right in thinking that is not even "clean room" in the way people usually think of it, e.g. Compaq?

The "clean room" aspect for that came in the way that the people writing the new implementation had no knowledge of the original source material, they were just given a specification to implement (see also Oracle v. Google).

If you're feeding an LLM GPL'd code and it "creates" something "new" from it, that's not "clean room", right?

At the end of the day the supposed reimplementation that the LLM generates isn't copyrightable either so maybe this is all moot.

> If you're feeding an LLM GPL'd code and it "creates" something "new" from it, that's not "clean room", right?

I didn’t RTFA but I suppose that by clean room here they mean you feed the code to ”one” LLM and tell it to write a specification. Then you give the specification to ”another” LLM and tell it to implement the specification.