Waiting for Terry Tao's thoughts, but these kind of things are good use of AI. We need to make science progress faster rather than disrupting our economy without being ready.
Definitely interesting.
Two thoughts. First, are the IMO questions somewhat related to other openly available questions online, making it easier for LLMs that are more efficient and better at reasoning to deduce the results from the available content?
Second, happy to test it on open math conjectures or by attempting to reprove recent math results.
From that thread: "The model solved P1 through P5; it did not produce a solution for P6."
It's interesting that it didn't solve the problem that was by far the hardest for humans too. China, the #1 team got only 21/42 points on it. In most other teams nobody solved it.
Wow. That's an impressive result, but how did they do it?
Wei references scaling up test-time compute, so I have to assume they threw a boatload of money at this. I've heard talk of running models in parallel and comparing results - if OpenAI ran this 10000 times in parallel and cherry-picked the best one, this is a lot less exciting.
If this is legit, then we need to know what tools were used and how the model used them. I'd bet those are the 'techniques to make them better at hard to verify tasks'.
I kid, this is actually pretty amazing!! I've noticed over the last several months that I've had to correct it less and less when dealing with advanced math topics so this aligns.
This is an awesome progress in human achievement to get these machines intelligent. And this is also a fast regress and decline on the human wisdom!
We are simply greasing the grooves and letting things slide faster and faster and calling it progress. How does this help to make the human and nature integration better?
Does this improve climate or make humans adapt better to changing climate? Are the intelligent machines a burning need for the humanity today? Or is it all about business and political dominance? At what cost? What's the fall out of all this?
I encourage anyone who thinks these are easy high-school problems to try to solve some. They're published (including this year's) at https://www.imo-official.org/problems.aspx. They make my head spin.
I am neither an optimist nor a pessimist for AI. I would likely be called both by the opposite parties. But the fact that AI / LLM is still rapidly improving is impressive in itself and worth celebrating for. Is it perfect, AGI, ASI? No. Is it useless? Absolutely not.
I am just happy the prize is so big for AI that there are enough money involve to push for all the hardware advancement. Foundry, Packaging, Interconnect, Network etc, all the hardware research and tech improvements previously thought were too expensive are now in the "Shut up and take my money" scenario.
Am I missing something or is this completely meaningless? It's 100% opaque, no details whatsoever and no transparency or reproducibility.
I wouldn't trust these results as it is. Considering that there are trillions of dollars on the line as a reward for hyping up LLMs, I trust it even less.
In the RLHF sphere you could tell some AI company/companies were targeting this because of how many IMO RLHF’ers they were hiring specifically. I don’t think it’s really easy to say how much “progress” this is given that.
They were hiring IMO winners because IMO winners tend to be good at working on AI, not because they had the people specifically to make the AI better at math.
Uh no. I’m a math RLHF’er. When I get hired, I work on math/logic up to masters level because that’s my qualifications. Masters and PHD work on masters and PHD level. And IMO work on IMO math.
Every skill and skill level is specifically assigned and hired in the RLHF world.
Sometime the skill levels are fuzzier, but that’s usually very temporary.
And as been said already, IMO is a specific skill that even PHD math holders aren’t universally trained for.
In fact no car company claims “gold medal” performance in Olympic running even they can do that 100 yeas ago. Obviously since IMO does not generate much money so it is an easy target.
BTW; “Gold medal performance “ looks a promotional term for me.
Makes sense. Mathematicians use intuiton a lot to drive their solution seeking, and I suppose an AI such as an LLM could develop intuition too. Of course where AI really wins is search speed and the fact that an LLM really doesn't get tired when exploring different strategies and steps within each strategy.
However, I expect that geometric intuition may still be lacking mostly because of the difficulty of encoding it in a form which an LLM can easily work with. After all, Chatgpt still can't draw a unicorn [1] although it seems to be getting closer.
I think equally impressive is the performance of the OpenAI team at the "AtCoder World Tour Finals 2025" a couple of days ago. There were 12 human participants and only one did better than OpenAI.
I believe this company used to present its results and approach in academic papers with enough details so that it could be reproduced by third parties.
> this isn’t an IMO-specific model. It’s a reasoning LLM that incorporates new experimental general-purpose techniques.
> it’s also more efficient [than o1 or o3] with its thinking. And there’s a lot of room to push the test-time compute and efficiency further.
> As fast as recent AI progress has been, I fully expect the trend to continue. Importantly, I think we’re close to AI substantially contributing to scientific discovery.
I thought progress might be slowing down, but this is clear evidence to the contrary. Not the result itself, but the claims that it is a fully general model and has a clear path to improved efficiency.
85 comments
[ 5.1 ms ] story [ 73.2 ms ] threadany details?
https://matharena.ai/imo/
Waiting for Terry Tao's thoughts, but these kind of things are good use of AI. We need to make science progress faster rather than disrupting our economy without being ready.
In 2021 Paul Christiano wrote he would update from 30% to "50% chance of hard takeoff" if we saw an IMO gold by 2025.
He thought there was an 8% chance of this happening.
Eliezer Yudkowsky said "at least 16%".
Source:
https://www.lesswrong.com/posts/sWLLdG6DWJEy3CH7n/imo-challe...
Second, happy to test it on open math conjectures or by attempting to reprove recent math results.
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.
It's interesting that it didn't solve the problem that was by far the hardest for humans too. China, the #1 team got only 21/42 points on it. In most other teams nobody solved it.
Wei references scaling up test-time compute, so I have to assume they threw a boatload of money at this. I've heard talk of running models in parallel and comparing results - if OpenAI ran this 10000 times in parallel and cherry-picked the best one, this is a lot less exciting.
If this is legit, then we need to know what tools were used and how the model used them. I'd bet those are the 'techniques to make them better at hard to verify tasks'.
I kid, this is actually pretty amazing!! I've noticed over the last several months that I've had to correct it less and less when dealing with advanced math topics so this aligns.
We are simply greasing the grooves and letting things slide faster and faster and calling it progress. How does this help to make the human and nature integration better?
Does this improve climate or make humans adapt better to changing climate? Are the intelligent machines a burning need for the humanity today? Or is it all about business and political dominance? At what cost? What's the fall out of all this?
I am just happy the prize is so big for AI that there are enough money involve to push for all the hardware advancement. Foundry, Packaging, Interconnect, Network etc, all the hardware research and tech improvements previously thought were too expensive are now in the "Shut up and take my money" scenario.
I wouldn't trust these results as it is. Considering that there are trillions of dollars on the line as a reward for hyping up LLMs, I trust it even less.
Every skill and skill level is specifically assigned and hired in the RLHF world.
Sometime the skill levels are fuzzier, but that’s usually very temporary.
And as been said already, IMO is a specific skill that even PHD math holders aren’t universally trained for.
BTW; “Gold medal performance “ looks a promotional term for me.
However, I expect that geometric intuition may still be lacking mostly because of the difficulty of encoding it in a form which an LLM can easily work with. After all, Chatgpt still can't draw a unicorn [1] although it seems to be getting closer.
[1] https://gpt-unicorn.adamkdean.co.uk/
Not sure there is a good writeup about it yet but here is the livestream: https://www.youtube.com/live/TG3ChQH61vE.
Now it is just doing a bunch of tweets?
> this isn’t an IMO-specific model. It’s a reasoning LLM that incorporates new experimental general-purpose techniques.
> it’s also more efficient [than o1 or o3] with its thinking. And there’s a lot of room to push the test-time compute and efficiency further.
> As fast as recent AI progress has been, I fully expect the trend to continue. Importantly, I think we’re close to AI substantially contributing to scientific discovery.
I thought progress might be slowing down, but this is clear evidence to the contrary. Not the result itself, but the claims that it is a fully general model and has a clear path to improved efficiency.
https://x.com/polynoamial/status/1946478249187377206