>In the end, the information in the training texts is a distilled proxy for the world
This is routinely asserted. How has it been proven?
Humans write all sorts of text that has zero connection to reality, even when they are ostensibly writing about reality.
Training on ancient greek philosophy which was expressly written to distill knowledge about the real world would produce a stupid LLM that doesn't know about the real world, because the training text was itself wrong about the underlying world.
Also, if LLMs were able to extract underlying truth from training material, why can't they do math very well? It would be easy to train an LLM on only correct math, and indeed you could generate any size corpus of provably correct math you want. I assume someone somewhere has demonstrated success training a neural network on math and having it regenerate something like "addition" or whatever, but how well would such a process survive if a large fraction of it's training material was instead just incorrect math?
The training text is nothing more than human generated text, and asserting anything about that more concrete than "Humans consider this text good enough to be worth writing" is fallacious.
This even applies if your training corpus is, for example, only physics scientific papers that have been strongly replicated and are likely "true". Unless the LLM is also trained on the data itself, the only information available is what the humans thought and wrote. There's no definite link between that and actual reality, which is why physics accepted an "Aether" for so long. The data we had up to that point aligned with our incorrect models. You could not disambiguate between the wrong Aetheric models and a better model with the data we had, and that would remain true of text written about the data.
Humans suck at distilling fact out of reality despite our direct connection to it for all sorts of fun reasons you can read about in psychology, but if you disconnect a human from reality, it only gets worse.
Why would you believe LLMs could possibly be different? A model trained on bad data cannot magically figure out which data is bad.
