Ask HN: Could AI's RH proof have used human work found in its training data?
When reading of third-party validations of AI's RH proof [0] that:
> While still possible, the proof exploiting a bug is vanishingly small now.
it makes me think we're realising in real-time that we're living through the most important time in human history. Moreso than during the invention of computers, the internet, or the wheel.
Problem: many people worked on RH, and some would have stored their attempts at proofs in places like GitHub or google docs - places that may have been ended up in training datasets.
Therefore, is it possible that Claude knew some significant parts of the proof (potentially even all of it but scattered across multiple private sources) and it merely assembled the parts into a single proof, as opposed to having thought of the proof for itself? [1]
(The distinction matters because assembling parts of existing proofs seems different to authoring a proof by oneself.)
[0] https://news.ycombinator.com/item?id=50590943
[1] I recall AI did something similar with an early Ergos problem (although it used public works), where it found connections and technically solved some unsolved problem, but Terence Tao came out a few days later saying it connected dots but didn't really think of anything truely novel.