in addition to the maths-by-LLM thing, I think a significant and overlooked component is just that LLMs are “good at” problem spaces that look like fuzzing. fuzzing is really effective when the scale is unfathomable and you can use (gasp) a real program to sift the success cases from the failures. it remains funny to me that the thing doing the actual work here is the checker program and not the supposedly magic LLM.