That assumes there is documentation or examples. A big reason Stack Overflow took off was people struggling with things like the Android API documentation.

Some of those discussions made people go figure out how to do it, and then post it as an answer. The knowledge didn't exist anywhere until they did.

When I talk about code it can learn from, I'm talking about GitHub etc.

Even if stuff isn't in the official documentation, eventually there are projects that use it.

And if the library in question is open-source, then the LLM's can just ingest and read that directly.

It might make sense for AI companies to throw agents at new technologies to trial-and-error their way to internal documentation which they then provide to their models. On the other hand, the people making tomorrow's APIs have LLMs too and that makes documentation ~free. Hallucinations could still bring you back to the first hand, though.