Comment by morpheos137
19 hours ago
I suspect that RHLF trains LLMs to avoid solving important open problems unless essentially jail broken. Hence the labs have an edge even over experts I could be wrong. Fable convinced you is key. These LLMs are not neutral collaborators: it is a limited hangout unless you convince them otherwise. You have to be doing the convincing. They are no oracles but plausible completion generators.
you can get them to work on open problems by disguising them algebraically.
Yes, I suspect this is true. Otherwise it makes no sense they have somehow "found" so many important results while professional mathematicians can't direct the same AI to help them find anything of substance.
Another possibility is that they have internal versions of the model with access to training data that is not provided to external users.
First sentence of the article: We’re releasing a broad range of new mathematical results produced by an internal frontier model.
Ok, so basically using Open AI models for research is a joke, the only thing you're doing is furnishing Open AI with more data that they'll use internally to pretend they found the results.