OpenAI用130B token解决纳维-斯托克斯猜想引发争议

Some more thoughts re Navier-Stokes by OpenAI

• Not entirely unexpected as this was supposed to be the Millennium Prize problem the most susceptible to finding a counterexample with known methods, but still, unexpectedly early. I would bet on 2027 for a problem of this calibre to be solved.
• I'm still pretty much shocked by it. Being a mathematician myself, Millennium Prize problems were always a pinnacle. There are of course more encompassing conjectures (Langlands programme), but these are the single most known problems out there.
• Funnily, I'm not sure it's AGI. Or we already had AGI with GPT-5.5 and Fable, because this is more of the same. 10,000 agents, 130b tokens, scaling test-time compute is the way to go, but isn't really novel nor very AGI-like.
• There's a very real cost to this result in millions of dollars of human time-compute. There are not many businesses that could throw compute at single problems: building rockets or drug research or hedge funds perhaps.
• I'm still pretty conflicted about this collective of narrow super-intelligences aka modern LLMs with harnesses. Are there diminishing returns past certain intelligence point? Is it all about the problem choice now to maximize your ROI? There's a reason we see less economic impact of AI than expected from this level of math breakthroughs. Generalization still seems to be the main missing skill that's hard to RL.