On: More questions about whether researchers can trust OpenAI with unpublished math

[auntienomen]

And conceptually novel approaches to outstanding problems are the sort of thing that a retrain should pick up on, because they would be hard to compress into what it already knows.