> If you feed it the entire dataset before 1905, LLMs aren't going to come up with general relativity.

Link?

You don't need a source for that, an LLM with such little data is barely able to form proper sentences.

> an LLM with such little data

There is a mountain of data pre-1905. Certainly enough to train a decent 30B parameter model.

Now, digitizing & OCRing all of that data... THAT is a challenge.