E.g here: https://pbs.twimg.com/media/GwLtrPeWIAUMDYI.png?name=orig
Frankly it looks to me like it's using an AlphaProof style system, going between natural language and Lean/etc. Of course OpenAI will not tell us any of this.
Anyway, that doesn't refute my point, it's just PR from a weaselly and dishonest company. I didn't say it was "IMO-specific" but the output strongly suggests specialized tooling and training, and they said this was an experimental LLM that wouldn't be released. I strongly suspect they basically attached their version of AlphaProof to ChatGPT.
We’re talking about Sam Altman’s company here. The same company that started out as a non profit claiming they wanted to better the world.
Suggesting they should be given the benefit of the doubt is dishonest at this point.
This is why HN threads about AI have become exhausting to read
Those models seem to be special and not part of their normal product line, as is pointed out in the comments here. I would assume that in that case they indeed had the purpose of passing these tests in mind when creating them. Or was it created for something different, and completely by chance they discovered they could be used for the challenge, unintentionally?
The bigger question is - why should everyone be excited by this? If they don't plan to share anything related to this AI model back to humanity.
The thing with IMO, is the solutions are already known by someone.
So suppose the model got the solutions beforehand, and fed them into the training model. Would that be an acceptable level of "cheating" in your view?
Finally, even if you aligned the model with the answers its weight shift of such an enormous model would be inconsequential. You would need to prime the context or boost the weights. All this seems like absurd lengths to go to to cheat on this one thing rather than focusing your energies on actually improving model performance. The payout for OpenAI isn’t a gold medal in the IMO it’s having a model that can get a gold medal at IMO then selling it. But it has to actually be capable of doing what’s on the tin otherwise their customers will easily and rapidly discover this.
Sorry, I like tin foil as much as anyone else, but this doesn’t seem credibly likely given the incentive structure.
> 5/N Besides the result itself, I am excited about our approach: We reach this capability level not via narrow, task-specific methodology, but by breaking new ground in general-purpose reinforcement learning and test-time compute scaling.
> 8/N Btw, we are releasing GPT-5 soon, and we’re excited for you to try it. But just to be clear: the IMO gold LLM is an experimental research model.
And from Sam Altman:
> we are releasing GPT-5 soon but want to set accurate expectations: this is an experimental model that incorporates new research techniques we will use in future models.
The wording you quoted is very tricky: the method used to create the model is generalizable, but the model is not a general-use model.
If I have a post-training method that allows a model excel at a narrow task, it's still a generalizable method if there's a wide range of narrow tasks that it works on.
https://x.com/alexwei_/status/1946477745627934979?s=46&t=Hov...
So its a big difference if you use a general intelligence system and makes it do well in math, or when you create a specialized system that is only good at math and can't be used to get good in other areas.
I agree with you. However, would a lot of working mathematicians score gold level without the IMO time constraints? Working mathematicians generally are not trying to solve a problem in the time span of one hour. I would argue that most working mathematicians, if given an arbitrary IMO problem and allowed to work on it for a week, would solve it. As for "gold level", with IMO problems you either solve one or you don't.
You could counter that it is meaningless to remove the time constraints. But we are comparing humans with OpenAI here. It is very likely OpenAI solved the IMO problems in a matter of minutes, maybe even seconds. When we talk about a chatbot achieving human-level performance, it's understood that the time is not a constraint on the human side. We are only concerned with the quality of the human output. For example: can OpenAI write a novel at the level of Jane Austen? Maybe it can, maybe it can't (for now) but Jane Austen was spending years to write such a novel, while our expectation is for OpenAI to do it at the speed of multiple words per second.
Addendum: Actually I am not sure the probability of solving it in a week is not much better than 6 hours for these questions because they are kind of random questions. But I agree with some parts of your post tbf.
Really? My expectation would have been the opposite, that time was a constraint for the AIs. OpenAI's highest end public reasoning models are slow, and there's only so much that you can do by parallelization.
Understanding how they dealt with time actually seems like the most important thing to put these results into context, and they said nothing about it. Like, I'd hope they gave the same total time allocation for a whole problem set as the human competitors. But how did they split that time? Did they work on multiple problems in parallel?
I did participate in math/informatics olympiads as a teenager and even taught it a little and from my experience, some type of people just _like_ that sort of problems naturally, they tickle their minds, and given time this people would develop to insane levels at it.
'Normal people', in my experience, even in math departments, don't like that type of problems, and would not fare well with them.
I grew up in a relatively underserved rural city. I skipped multiple grades in math, completed the first two years of college math classes while in high school, and won the award for being the best at math out of everyone in my school.
I've met and worked with a few IMO gold medalists. Even though I was used to scoring in the 99th percentile on all my tests, it felt like these people were simply in another league above me.
I'm not trying to toot my own horn. I'm definitely not that smart. But it's just ridiculous to shoot down the capabilities of these models at this point.
(Not to take away from the result, which I'm really impressed by!)
https://deepmind.google/discover/blog/alphago-zero-starting-...
They are all outstanding mathematicians, but the IMO type questions are not something that mathematicians can universally solve without preparation.
There are of course some places that pride themselves on only taking “high scoring” mathematicians, and people will introduce themselves with their name and what they scored on the Putnam exam. I don’t like being around those places or people.
My second degree is in mathematics. Not only can I probably not do these but they likely aren’t useful to my work so I don’t actually care.
I’m not sure an LLM could replace the mathematical side of my work (modelling). Mostly because it’s applied and people don’t know what they are asking for, what is possible or how to do it and all the problems turn out to be quite simple really.
So IMO is basically the leetcode of Mathematics.
"Did you win the Putnam?"
How do you know?
> This is not a model specialized to IMO problems.
Any proof?
There are trillions of dollars at stake in hyping up these products; I take everything these companies write with a cartload of salt.
From the thread:
> just to be clear: the IMO gold LLM is an experimental research model.
The thread tried to muddy the narrative by saying the methodology can generalize, but no one is claiming the actual model is a generalized model.
There'd be a massively different conversation needed if a generalized model that could become the next iteration of ChatGPT had achieved this level of performance.
Basically how you do RL is that you make a set of training examples of input-output pairs, and set aside a smaller validation set, which you never train on, to check if your model's doing well.
What you do is you tweak the architecture and the training set until it does well on the validation set. By doing so, you inadvertedly leak info about the training set. Perhaps you choose an architecture which does well on the validation set. Perhaps you train more on examples more like ones being validated.
Even without the explicit intent to cheat, it's very hard to avoid this contamination, if you chose a different validation set, you'd end up with a different model.
Professional mathematicians would not get this level of performance, unless they have a background in IMO themselves.
This doesn’t mean that the model is better than them in math, just that mathematicians specialize in extending the frontier of math.
The answers are not in the training data.
This is not a model specialized to IMO problems.