did o3 not still get the highest score so far?

@Bayesian Sam Altman had made an informal claim in February that their internal model had reached 50th in the world, which at the time implied a rating of 3045. I don't think that was ever actually published though.

@TimothyJohnson5c16 hmmmmmm yeah that should probably count?
does anyone have a case for it not counting

@Bayesian I guess the main case for not counting it would be if you think Sam Altman isn't trustworthy enough.

@TimothyJohnson5c16 right. he clearly isn't "generally" trustworthy but this is the kind of thing he is unlikely to make up

AI got 2nd on AtCoder Grand Finals Heuristics contents (1 task, 10 hours) and got gold on IMO (6 tasks, 9 hours). Codeforces rounds have more top participants (AtCoders was just 12 people, IMO is restricted by age) but format of 5-7 tasks and 2-3 hours favors AI.

Although tbh due to the current state of the leaderboard, 4000 would need AI to dominate humans, so maybe I shouldn't have bet YES on 4000+ (well it's too late)

same

4000+ at 24% is kinda wild. 1300 pt leap in 6 months?

tbf I'm not familiar with codeforces. But yeah the best human is 3700.

top 150 (they claimed top 50 internally) to significantly better than the best human is wild.

@ChinmayTheMathGuy I guess internally they're around 3100 or so now

@ChinmayTheMathGuy humans have been 4000+ several times by now, but the top ratings on codeforces are extremely volatile; if you're the top rated coder you have to win every contest to avoid bleeding massive amounts of rating

@zsig if im not mistaken only 2 people have been 4000+ ?

by score, do you mean a score on a contest or an overall rating?

The setting analogous to the o3 case. Do you know if that was a contest or from overall rating? as far as I can tell it seems to be overall rating, in which case this market will deal with overall rating as well

@Bayesian thanks ! Somehow I seriously doubt that an ai will get 4000+ Thats the equivalent of ai beating humans almost completely in competitive programming.

@Bayesian indeed, the o3 case was talking about overrall rating.

@ZandaZhu I also doubt it tbh but not 25%, ig higher. openai was eyeing #1 at competitive programming by EOY, a bit after the o3 release

@Bayesian NO limit order at 30% for you.

@TimothyJohnson5c16 I agree, I think ai will not surpass humans at competitive programming as early as 2025, if this really happens us humans are closer to being doomed than we might think.

Also this question should at least be somewhat constrained by compute. The o3 rating was done by ranking 1062 prompts and also with 10 passes.

So yes with infinite compute I don't doubt it can achieve 4k but that's not the point of the question imo. If you took 1000 humans they could also achieve 5000+ rating.

@patrik I don’t think if you took 100 humans they could achieve 5000+ rating!

@Bayesian The top 100 for sure. Getting high ratin in codeforces is only about consistently beating everyone.

@patrik Sorry i typoed i meant to write 1000 humans. And yeah i thought u were selecting at random 1000 humans. I agree

@Paul why?

https://x.com/tsarnick/status/1888111042301211084