Didn't top math researchers call AI progress absolutely real and dangerous for math? It's not just HN commenters that are impressed!

▲

ssfdg 2 hours ago | parent [-]

By all accounts the "dangerous for math" claims seem to be primarily around flooding the field with complicated impossible-to-understand proofs that according to recent research may or may not be correct depending on what's going on with the Lean implementation.

It's looking to me like it's more of a slop PR problem than it is that these things are genius at math and will displace mathematicians. I am happy to be wrong but I strongly suspect the next few weeks to months will result in more and more of this work being exposed as slop.

These things are ok-ish to halfway decent at coding tasks with a ton of babysitting and still make tons of extremely simple errors almost constantly, why should math be any different?

	▲	azan_ 37 minutes ago \| parent [-]
		Yes, Lean verified proofs could be wrong, but the chances for that are much smaller than human not spotting error (in absence of formal verification). IIRC the main concern that Tao voiced are indeed impenetrable proofs that humans won't understand, but not concerns about truthfulness (I might have missed something though, so if he or other Fields medalists have talked about that recently I'd be grateful if you could link it). > These things are ok-ish to halfway decent at coding tasks with a ton of babysitting and still make tons of extremely simple errors almost constantly, why should math be any different? 1) AI is winning programming competitions, 2025 was probably the last year we've had human participant winning* 2) Math is different because there's formal verification. * Of course competitive programming is different than enterprise programming, but competitive programming is closer to math.