Can someone tell me where your average every day human that’s walking around and has a regular job and kids and a mortgage would land on this leaderboard? That’s who we should be comparing against.
The fact that the only formal comparisons for AI systems that are ever done are explicitly based on the highest performing narrowly focused humans, tells me how unprepared society is for what’s happening.
Appreciate that: at the point in which there is unambiguous demonstration of superhuman level performance across all human tasks by a machine, (and make no mistake, that *is the bar that this blog post and every other post about AI sets*) it’s completely over for the human race; unless someone figures out an entirely new economic system.
The average person is bad at literally almost everything.
If I want something done, I'll seek out someone with a skill set that matches the problem.
I don't want AI to be as good as an average person. I want AI to be better than the person I would go to for help. A person can talk with me, understand where I've misunderstood my own problem, can point out faulty assumptions, and may even tell me that the problem isn't even a problem that needs solving. A person can suggest a variety of options and let me decide what trade-offs I want to make.
If I don't trust the AI to do that, then I'm not sure why I'd use it for anything other than things that don't need to be done at all, unless I can justify the chance that maybe it'll be done right, and I can afford the time lost getting it done right without the AI afterwards.
I really don't think it does, because we disagree on what the upper bound of an LLM is capable of reasoning about.
An average human may not be suitable for a given task, but a person with specialized skills will be. More than that, I believe they will continue to outperform LLMs on solving unbounded problems- i.e. those problems without an obvious, algorithmic solution.
Anything that requires brute force computation can be done by an LLM more quickly, assuming you have humans you trust to validate the output, but that's about the extent of what I'm expecting them to achieve.
The average person can’t play 99% of all musical instruments, speak 99% of all languages, do 99% of surgeries, recite 99% of all poems from memory etc.
We don’t ask the average person to do most things, either finding a specialist or providing training beforehand.
You and the parent poster seem to be conflating the ideas of:
- Does not have the requisite skills and experiences to do X successfully
- Inherently does not have the capacity to do X
I think the former is a reasonable standard to apply in this context. I'd definitely say I would be bad if I tried to play the guitar, but I'm not inherently incapable of doing it. It's just not very useful to say "I could be good at it if I put 1000 hours of practice in."
That's why there's the qualifier of "average person". If one learns to play the guitar well, they are no longer the average person in the context of guitar playing.
> Obviously you could train someone to recite the The Raven from memory, but they can’t do it now.
That doesn't make them bad at reciting The Raven from memory. Being trained to recite The Raven from memory and still being unable to do so would be a proper application of the term. There is an obvious difference between the two states of being and conflating them is specious.
If you want to take seriously the premise that humans are bad at almost everything because most humans haven't been trained at doing almost everything humans can do, then you must apply the same rubric to LLMs, which are only capable of expressions within their specific dataset (and thus not the entire corpus of data on which they haven't been trained) and even then which tend to confabulate far more frequently than human beings at even simple tasks.
edit: never mind, I guess you aren't willing to take this conversation on good faith.
Didn't this start with "Can someone tell me where your average every day human that’s walking around and has a regular job and kids and a mortgage would land on this leaderboard? That’s who we should be comparing against."
And the average person would do poorly. Not because they couldn't be trained to do it, but because they haven't.
It's obvious that the average person would do bad at the International Math Olympiad. Although I don't know why the qualifiers of "regular job and kids and a mortgage" are necessary, except as a weird classist signifier. I strongly suspect most people on HN, who consider themselves set apart from the average, with some also having a regular job, kids and a mortgage, would also not do well at the International Math Olympiad.
But that isn't the claim I'm objecting to. The claim I'm objecting to is "The average person is bad at literally almost everything," which is not an equivalent claim to "people who aren't trained at math would be bad at math at a competitive level," because it implicitly includes everything that a person is trained in and is expected to be qualified to do.
It was just bad, cynical hyperbole. And it's weird that people are defending it so aggressively.
It's obvious that 'bad at' in this context means 'incapable of doing well'.
Nitpicking language doesn't help to move the conversation. One thing most humans are good at is understanding meaning even when the speaker wasn't absolutely precise.
More than 50% of people cannot write a 'hello world' program in any programming language.
More than 50% of people employed as software engineers cannot read an academic paper in a field like education, and explain whether the conclusions are sound, based on the experiment description and included data.
More than 50% of people cannot interpret an X-ray.
> More than 50% of people employed as software engineers cannot read an academic paper in a field like education, and explain whether the conclusions are sound, based on the experiment description and included data.
I know this was meant as a dig, but I’m actually guessing that software engineers score higher on this task than non-engineers who hold M.Ed. degrees.
Agreed! Probably 3% of software could do it, vs 1% for M.Ed holders.
The only reason I chose software engineers is because I was trying to show that people who can write 'hello world' programs (first example) are not good at all intellectual tasks.
Machines have always had superhuman capabilities in narrow domains. The LLM domain is quite broad but it's still just a LLM, beholden to its training.
The average everyday human does not have the time to read all available math texts. LLMs do, but they still can't get bronze. What does that say about them?
One interesting takeaway for me, a non-practitioner, was that the models appears to be fairy decent at judging their own output.
They used best-of-32 and used the same model to judge a "tournament" to find the best answer. Seems like something that could be boltet on reasonably easy, eg in say WebUI.
edit: forgot to add that I'm curious if this translates to smaller models as well, or if it requires these huge models.
The whole competition is unfair anyway. An "AI" has access to millions of similar problems stolen and encoded in the model. Humans would at least need access to a similar database; think open database exam, a nuclear version of open book exam.
Easy benchmark that's hard to fake: data compression. Intelligence is largely about creating compact predictive models and so is data compression. The output should be a program generating the sequence or the dataset, based on entry id or nearby data points. Typical LLM bullshit won't work here because the output isn't English prose that can fool a human.
> For Problem 5, models often identified the correct strategies but failed to prove them, which is, ironically, the easier part for an IMO participant. This contrast ... suggests that models could improve significantly in the near future if these relatively minor logical issues are addressed.
Interesting but I'm not sure if this is really due to "minor logical issues". This sounds like a failure due to the lack of the actual understanding (the world model problem). Perhaps the actual answers from AIs might have some hints, but I can't find them.
(EDIT: ooops, found the output on the main page of their website. Didn't expect that.)
> Best-of-n is Important ... the models are surprisingly effective at identifying the relative quality of their own outputs during the best-of-n selection process and are able to look past coherence to check for accuracy.
Yes, monkeys could write Shakespeare works given enough time.
But in this case, it is really hard to know if a model is identifying "correct answers" reliably. A lot of answers are really hard to qualify as correct or not when written by humans, much more when written by a machine trying to trick readers into thinking the answer is correct. It can be done, but I doubt LLM are being trained to identify the subtle differences between those types of potential answers.
What else should people do? If we just saturate at "wow this is amazing!" there's nothing to talk about, nothing to evaluate, nothing to push the boundaries forward further (or caution against doing so).
Yes, we're all impressed, but it's time to move on and start looking at where the frontier is and who's on it.
LLMs are really good with words and kind of crap at “thinking.” Humans are wired to see these two things as tightly connected. A machine that thinks poorly and talks great is inherently confusing. A lot of discussion and disputes around LLMs comes down to this.
It wasn’t that long ago that the Turing Test was seen as the gold standard of whether a machine was actually intelligent. LLMs blew past that benchmark a year or two ago and people barely noticed. This might be moving the goalposts, but I see it as a realization that thought and language are less inherently connected than we thought.
So yeah, the fact that they even do this well is pretty amazing, but they sound like they should be doing so much better.
> LLMs are really good with words and kind of crap at “thinking.” Humans are wired to see these two things as tightly connected. A machine that thinks poorly and talks great is inherently confusing. A lot of discussion and disputes around LLMs comes down to this.
It's not an unfamiliar phenomenon in humans. Look at Malcolm Gladwell.
> Each model was run with the recommended hyperparameters and a maximum token limit of 64,000. No models needs more than this number of tokens
I'm a little confused by this. My assumptions (possibly incorrect!): 64k tokens per prompt, they are claiming the model wouldn't need more tokens even for reasoning
Is that right? Would be helpful to see how many tokens the models actually used.
99.99+% of all problems humans face do not require particularly original solutions. Determining whether LLMs can solve truly original (or at least obscure) problems is interesting, and a problem worth solving, but ignores the vast majority of the (near-term at least) impact they will have.
People who say this don’t understand the breakthrough we had in the last couple of years. 15 years ago I was laughing at people predicting AI would turn everything upside down soon. I’m not laughing anymore. I’ve been around long enough to see some AI hype cycles and this time it is different.
15 years ago I, working on AI systems at a FAANG, would have told you “real” AI probably wasn’t coming in my lifetime. 15 years ago the only engineers I knew who thought AI was coming soon were dreamers and Silicon Valley koolaiders. The rest of us saw we needed a step-function break through that may not even exist. But it did, and we got there, a couple of years ago.
Now I’m telling people it’s here. We’ve hit a completely different kind of technology, and it’s so clear to people working in the field. The earthquake has happened and the tsunami is coming.
the value of human beings isn't in their capacity to do routine tasks but to respond with some common sense to all the critical issues in the 2% at the tail.
This is why original problems are important, it's a measure of how sensible something is in an open-ended environment, and here they're completely useless, not just because they fail but how they fail. The fact that these LLMS according to the article "invent non-existent math theorems", i.e. gibberish instead of even being able to know what they don't know, is an indication of how limited this still is.
To be frank, I take precisely the opposite view. Most people solve novel problems every day, mostly without thinking much about it. Our inability to perceive the immense complexity of the things we do every day is merely due to familiarity. In other words we're blind to the details because our brain handles them automatically, not because they don't exist.
Software engineers understand this better than most - describing a task in general terms, and doing it yourself, can be incredibly easy, even while writing the code to automate the task is difficult or impossible, because of all the devilish details we don't often think about.
I work with developers every day. Between us we often give the AI directions like:
* Write a query to link table X to table Y across this schema, returning all the unique entries related to X.id 1234
* Write code add an editable comment list to this UI
* Give me a design to visually manage statuses for this list
* Look at this UI and give me five ideas for improving it
Some of those work better than others, but none of them are guaranteed failures.
I was hoping to see the questions (which I can probably find online), but also the answers from models and the judge's scores! Am I missing a link? Without that I can't tell whether I should be impressed or not.
> Gemini 2.5 Pro achieved the highest score with an average of 31% (13 points). While this may seem low, especially considering the $400 spent on generating just 24 answers
What? That’s some serious cash for mostly wrong answers.
In a few months (weeks, days - maybe it has already happened) models will have much better performance on this test.
Not because of actual increased “intelligence” but because the test would be included in model’s training data - either directly or indirectly where model developers “tune” their model to give better performance on this particular attention driving test.
What they mean is that in a couple of weeks there are going to be stories titled "LLMS NOW BETTER THAN HUMANS AT 2025 INTERNATIONAL MATH OLYMPIAD" (stories published as thinly-veiled investment solicitations) but in reality they're still shitty-- they've just had the answers fed in to be spit back out.
You can really only do a fair reproducible test if the models are static and not sitting behind an api where you have no idea on how they are updated or continuously tweaked.
This particular test is heralded as some sort of breakthrough and the companies in this field are raising billions of dollars from investors and paying their star employees tens of millions.
The economic incentives to tweak, tune, or cheat are through the roof.
I'd also be highly wary of the method they used because of statements like this:
>we note that the vast majority of its answers simply stated the final answer without additional justification
While the reasoning steps are obviously important for judging human participant answers, none of the current big-game providers disclose their actual reasoning tokens. So unless they got direct internal access to these models from the big companies (which seems highly unlikely), this might be yet another failed study designed to (of which we have seen several in recent months, even by serious parties).
Wow, that’s incredible. Cats are progressing so fast, especially unreleased cats seem to be doing much better. My two orange kitties aren’t doing well on math problems but obviously that’s because I’m not prompting the right way – any day now. If I ever get it to work, I’ll be sure to share the achievements on X, while carefully avoiding explaining how I did it or provide any data that can corroborate the claims.
That's a claim that is far less plausible. OpenAI could have thrown more resources at the problem and I would be surprised if that didn't improve the results.
The last time someone claimed a medal in an olympiad like this, turned out they manually translated the problem into Lean and then ran a brute force search algorithm to find a proof. For 60 hours. On a supercomputer.
Meanwhile high schoolers get a piece of paper and 4.5 hours.
Even though chess is now effectively solved against human players, I still remember Kasparov's suspicion that one of Deep Blue's moves had a human touch. It was never proven or disproven, but I trust Kasparov's deep intuition amplified by Kasparov requesting access to Deep Blue’s logs, and IBM refusing to share them in full. For more discussions see [1][2][3].
Its quite reasonable. We have yet to meet anything more intelligent than humans so why do we think we can create something more intelligent than us when we don't fully understand the complexities how we work?
AI still has a long way to go, though it has proven to be a useful tool at this point.
who said anything about creating something more intelligent than us, these articles have the air of "why are we wasting our time on this stuff", people like gary marcus link them, meanwhile they get better week over week
Can someone tell me where your average every day human that’s walking around and has a regular job and kids and a mortgage would land on this leaderboard? That’s who we should be comparing against.
The fact that the only formal comparisons for AI systems that are ever done are explicitly based on the highest performing narrowly focused humans, tells me how unprepared society is for what’s happening.
Appreciate that: at the point in which there is unambiguous demonstration of superhuman level performance across all human tasks by a machine, (and make no mistake, that *is the bar that this blog post and every other post about AI sets*) it’s completely over for the human race; unless someone figures out an entirely new economic system.
The average person is bad at literally almost everything.
If I want something done, I'll seek out someone with a skill set that matches the problem.
I don't want AI to be as good as an average person. I want AI to be better than the person I would go to for help. A person can talk with me, understand where I've misunderstood my own problem, can point out faulty assumptions, and may even tell me that the problem isn't even a problem that needs solving. A person can suggest a variety of options and let me decide what trade-offs I want to make.
If I don't trust the AI to do that, then I'm not sure why I'd use it for anything other than things that don't need to be done at all, unless I can justify the chance that maybe it'll be done right, and I can afford the time lost getting it done right without the AI afterwards.
Which proves my point precisely that unless you’re superhuman in this definition, you’re obsolete.
Nothing new really, but there’s no where left to go for human labor and even that concept is being jeered at as a fantasy despite this attitude.
I really don't think it does, because we disagree on what the upper bound of an LLM is capable of reasoning about.
An average human may not be suitable for a given task, but a person with specialized skills will be. More than that, I believe they will continue to outperform LLMs on solving unbounded problems- i.e. those problems without an obvious, algorithmic solution.
Anything that requires brute force computation can be done by an LLM more quickly, assuming you have humans you trust to validate the output, but that's about the extent of what I'm expecting them to achieve.
Think beyond LLMs
You need to think about what comes after LLMs that look nothing like LLMs
You need to think about what robots with human capabilities, which are improving multiple times per day, is going to do.
Now add LLMs back in as your HMI
"The average person is bad at literally almost everything."
Wow... that's quite a generalization. And not my experience at all.
The average person can’t play 99% of all musical instruments, speak 99% of all languages, do 99% of surgeries, recite 99% of all poems from memory etc.
We don’t ask the average person to do most things, either finding a specialist or providing training beforehand.
One cannot be bad at the things one doesn't even do. None of this demonstrates that humans are bad at "literally almost everything."
You and the parent poster seem to be conflating the ideas of:
- Does not have the requisite skills and experiences to do X successfully
- Inherently does not have the capacity to do X
I think the former is a reasonable standard to apply in this context. I'd definitely say I would be bad if I tried to play the guitar, but I'm not inherently incapable of doing it. It's just not very useful to say "I could be good at it if I put 1000 hours of practice in."
That's why there's the qualifier of "average person". If one learns to play the guitar well, they are no longer the average person in the context of guitar playing.
> One cannot be bad at the things one doesn't even do.
??? If you don’t know how to do something you’re really bad at it. I’m not sure what that sentence is even trying to convey.
> Obviously you could train someone to recite the The Raven from memory, but they can’t do it now.
That doesn't make them bad at reciting The Raven from memory. Being trained to recite The Raven from memory and still being unable to do so would be a proper application of the term. There is an obvious difference between the two states of being and conflating them is specious.
If you want to take seriously the premise that humans are bad at almost everything because most humans haven't been trained at doing almost everything humans can do, then you must apply the same rubric to LLMs, which are only capable of expressions within their specific dataset (and thus not the entire corpus of data on which they haven't been trained) and even then which tend to confabulate far more frequently than human beings at even simple tasks.
edit: never mind, I guess you aren't willing to take this conversation on good faith.
Didn't this start with "Can someone tell me where your average every day human that’s walking around and has a regular job and kids and a mortgage would land on this leaderboard? That’s who we should be comparing against."
And the average person would do poorly. Not because they couldn't be trained to do it, but because they haven't.
It's obvious that the average person would do bad at the International Math Olympiad. Although I don't know why the qualifiers of "regular job and kids and a mortgage" are necessary, except as a weird classist signifier. I strongly suspect most people on HN, who consider themselves set apart from the average, with some also having a regular job, kids and a mortgage, would also not do well at the International Math Olympiad.
But that isn't the claim I'm objecting to. The claim I'm objecting to is "The average person is bad at literally almost everything," which is not an equivalent claim to "people who aren't trained at math would be bad at math at a competitive level," because it implicitly includes everything that a person is trained in and is expected to be qualified to do.
It was just bad, cynical hyperbole. And it's weird that people are defending it so aggressively.
It's obvious that 'bad at' in this context means 'incapable of doing well'.
Nitpicking language doesn't help to move the conversation. One thing most humans are good at is understanding meaning even when the speaker wasn't absolutely precise.
More than 50% of people cannot write a 'hello world' program in any programming language.
More than 50% of people employed as software engineers cannot read an academic paper in a field like education, and explain whether the conclusions are sound, based on the experiment description and included data.
More than 50% of people cannot interpret an X-ray.
> More than 50% of people employed as software engineers cannot read an academic paper in a field like education, and explain whether the conclusions are sound, based on the experiment description and included data.
I know this was meant as a dig, but I’m actually guessing that software engineers score higher on this task than non-engineers who hold M.Ed. degrees.
Agreed! Probably 3% of software could do it, vs 1% for M.Ed holders.
The only reason I chose software engineers is because I was trying to show that people who can write 'hello world' programs (first example) are not good at all intellectual tasks.
Machines have always had superhuman capabilities in narrow domains. The LLM domain is quite broad but it's still just a LLM, beholden to its training.
The average everyday human does not have the time to read all available math texts. LLMs do, but they still can't get bronze. What does that say about them?
Average humans, no. Mathematicians with enough time and a well indexed database of millions of similar problems, probably.
We don't allow chess players to access a Syzygy tablebase in a tournament.
Average human would score exactly 0 at IMO.
That’s not how modern societies/economies work.
We have specialists everywhere.
My literal last sentence addresses this
> average every day human
Average math major can't get Brozne.
One interesting takeaway for me, a non-practitioner, was that the models appears to be fairy decent at judging their own output.
They used best-of-32 and used the same model to judge a "tournament" to find the best answer. Seems like something that could be boltet on reasonably easy, eg in say WebUI.
edit: forgot to add that I'm curious if this translates to smaller models as well, or if it requires these huge models.
So the gold medal claims in https://news.ycombinator.com/item?id=44613840 look exaggerated.
The whole competition is unfair anyway. An "AI" has access to millions of similar problems stolen and encoded in the model. Humans would at least need access to a similar database; think open database exam, a nuclear version of open book exam.
Easy benchmark that's hard to fake: data compression. Intelligence is largely about creating compact predictive models and so is data compression. The output should be a program generating the sequence or the dataset, based on entry id or nearby data points. Typical LLM bullshit won't work here because the output isn't English prose that can fool a human.
> For Problem 5, models often identified the correct strategies but failed to prove them, which is, ironically, the easier part for an IMO participant. This contrast ... suggests that models could improve significantly in the near future if these relatively minor logical issues are addressed.
Interesting but I'm not sure if this is really due to "minor logical issues". This sounds like a failure due to the lack of the actual understanding (the world model problem). Perhaps the actual answers from AIs might have some hints, but I can't find them.
(EDIT: ooops, found the output on the main page of their website. Didn't expect that.)
> Best-of-n is Important ... the models are surprisingly effective at identifying the relative quality of their own outputs during the best-of-n selection process and are able to look past coherence to check for accuracy.
Yes, it's always easier to be a backseat driver.
>Yes, it's always easier to be a backseat driver
Any model that can identify the correct answer reliably can arrive at the correct answer given enough time and stochasticity.
Yes, monkeys could write Shakespeare works given enough time.
But in this case, it is really hard to know if a model is identifying "correct answers" reliably. A lot of answers are really hard to qualify as correct or not when written by humans, much more when written by a machine trying to trick readers into thinking the answer is correct. It can be done, but I doubt LLM are being trained to identify the subtle differences between those types of potential answers.
NP
How quickly we shift our expectations. If you told me 5 years ago we'd have technology that can do this, I wouldn't believe you.
This isn't to say we shouldn't think critically about the use and performance of models, but "Not Even Bronze..." turned me off to this critique.
What else should people do? If we just saturate at "wow this is amazing!" there's nothing to talk about, nothing to evaluate, nothing to push the boundaries forward further (or caution against doing so).
Yes, we're all impressed, but it's time to move on and start looking at where the frontier is and who's on it.
In 2024 AlphaProof got Silver level, so people righteously expect a lot now.
(It's specifically trained on formalized math problems, unlike most LLM, so it's not an apple to apple comparison.)
LLMs are really good with words and kind of crap at “thinking.” Humans are wired to see these two things as tightly connected. A machine that thinks poorly and talks great is inherently confusing. A lot of discussion and disputes around LLMs comes down to this.
It wasn’t that long ago that the Turing Test was seen as the gold standard of whether a machine was actually intelligent. LLMs blew past that benchmark a year or two ago and people barely noticed. This might be moving the goalposts, but I see it as a realization that thought and language are less inherently connected than we thought.
So yeah, the fact that they even do this well is pretty amazing, but they sound like they should be doing so much better.
> LLMs are really good with words and kind of crap at “thinking.” Humans are wired to see these two things as tightly connected. A machine that thinks poorly and talks great is inherently confusing. A lot of discussion and disputes around LLMs comes down to this.
It's not an unfamiliar phenomenon in humans. Look at Malcolm Gladwell.
Here are the IMO problems if you want to give them a try:
https://www.imo-official.org/year_info.aspx?year=2025 (download page)
They are very difficult.
> Each model was run with the recommended hyperparameters and a maximum token limit of 64,000. No models needs more than this number of tokens
I'm a little confused by this. My assumptions (possibly incorrect!): 64k tokens per prompt, they are claiming the model wouldn't need more tokens even for reasoning
Is that right? Would be helpful to see how many tokens the models actually used.
they didn't even do a (non-ml) agentic descent? like have a quicky api that requeries itself generating new context?
"ok here is my strategy here are the five steps", then requery with a strategy or proof of step 1, 2, 3...
in a dfs
99.99+% of all problems humans face do not require particularly original solutions. Determining whether LLMs can solve truly original (or at least obscure) problems is interesting, and a problem worth solving, but ignores the vast majority of the (near-term at least) impact they will have.
15 years ago they were predicting that AI would turn everything upside down in 15 years time. It hasn't.
People who say this don’t understand the breakthrough we had in the last couple of years. 15 years ago I was laughing at people predicting AI would turn everything upside down soon. I’m not laughing anymore. I’ve been around long enough to see some AI hype cycles and this time it is different.
15 years ago I, working on AI systems at a FAANG, would have told you “real” AI probably wasn’t coming in my lifetime. 15 years ago the only engineers I knew who thought AI was coming soon were dreamers and Silicon Valley koolaiders. The rest of us saw we needed a step-function break through that may not even exist. But it did, and we got there, a couple of years ago.
Now I’m telling people it’s here. We’ve hit a completely different kind of technology, and it’s so clear to people working in the field. The earthquake has happened and the tsunami is coming.
Thank you for sharing your experience. It makes the impact of the recent advances palpable.
the value of human beings isn't in their capacity to do routine tasks but to respond with some common sense to all the critical issues in the 2% at the tail.
This is why original problems are important, it's a measure of how sensible something is in an open-ended environment, and here they're completely useless, not just because they fail but how they fail. The fact that these LLMS according to the article "invent non-existent math theorems", i.e. gibberish instead of even being able to know what they don't know, is an indication of how limited this still is.
To be frank, I take precisely the opposite view. Most people solve novel problems every day, mostly without thinking much about it. Our inability to perceive the immense complexity of the things we do every day is merely due to familiarity. In other words we're blind to the details because our brain handles them automatically, not because they don't exist.
Software engineers understand this better than most - describing a task in general terms, and doing it yourself, can be incredibly easy, even while writing the code to automate the task is difficult or impossible, because of all the devilish details we don't often think about.
I work with developers every day. Between us we often give the AI directions like:
Some of those work better than others, but none of them are guaranteed failures.I really doubt a contest for high schoolers contains any truly original problems.
"or at least obscure"
I was hoping to see the questions (which I can probably find online), but also the answers from models and the judge's scores! Am I missing a link? Without that I can't tell whether I should be impressed or not.
https://matharena.ai/
On their website you can see the full answers LLM gave ("click cells to see...")
> Gemini 2.5 Pro achieved the highest score with an average of 31% (13 points). While this may seem low, especially considering the $400 spent on generating just 24 answers
What? That’s some serious cash for mostly wrong answers.
The time investment a human has to make to get 31% on the IMO is worth far more than $400
The human still has to put in that time. How would you know what 31% is correct?
this makes me really wonder about what is the underlying practical mathematical skill?
intuition????
plus a little of skills
In a few months (weeks, days - maybe it has already happened) models will have much better performance on this test.
Not because of actual increased “intelligence” but because the test would be included in model’s training data - either directly or indirectly where model developers “tune” their model to give better performance on this particular attention driving test.
From the post: "Evaluation began immediately after the 2025 IMO problems were released to prevent contamination."
Doe this address your concern?
What they mean is that in a couple of weeks there are going to be stories titled "LLMS NOW BETTER THAN HUMANS AT 2025 INTERNATIONAL MATH OLYMPIAD" (stories published as thinly-veiled investment solicitations) but in reality they're still shitty-- they've just had the answers fed in to be spit back out.
Companies would game metrics whenever they have the opportunity. What else is new?
I suppose what's new is that the models aren't as smart as their companies claimed.
Not really.
Luckily there’s a new set of problems every year
You can really only do a fair reproducible test if the models are static and not sitting behind an api where you have no idea on how they are updated or continuously tweaked.
This particular test is heralded as some sort of breakthrough and the companies in this field are raising billions of dollars from investors and paying their star employees tens of millions.
The economic incentives to tweak, tune, or cheat are through the roof.
Related: https://news.ycombinator.com/item?id=44613840
Note that it's two different things:
This OP claims the publicly available models all failed to get Bronze.
OpenAI tweet claims there is an unreleased model that can get Gold.
I'd also be highly wary of the method they used because of statements like this:
>we note that the vast majority of its answers simply stated the final answer without additional justification
While the reasoning steps are obviously important for judging human participant answers, none of the current big-game providers disclose their actual reasoning tokens. So unless they got direct internal access to these models from the big companies (which seems highly unlikely), this might be yet another failed study designed to (of which we have seen several in recent months, even by serious parties).
My (unreleased) cat did even better than the OpenAI model. No you cannot see. Yes you have to trust me. Now gimme more money.
Wow, that’s incredible. Cats are progressing so fast, especially unreleased cats seem to be doing much better. My two orange kitties aren’t doing well on math problems but obviously that’s because I’m not prompting the right way – any day now. If I ever get it to work, I’ll be sure to share the achievements on X, while carefully avoiding explaining how I did it or provide any data that can corroborate the claims.
I don't know the details (of course, it's unreleased), but note that MathArena evaluated "average of 4 attempts", and limited token usages to 64k.
OpenAI likely had unlimited tokens, and evaluated "best of N attempts."
That's a claim that is far less plausible. OpenAI could have thrown more resources at the problem and I would be surprised if that didn't improve the results.
The model did not fit in the margin.
We'll never know how many GPUs and other assistance (like custom code paths) this model got.
Exactly. Whom to believe?
The last time someone claimed a medal in an olympiad like this, turned out they manually translated the problem into Lean and then ran a brute force search algorithm to find a proof. For 60 hours. On a supercomputer.
Meanwhile high schoolers get a piece of paper and 4.5 hours.
Even though chess is now effectively solved against human players, I still remember Kasparov's suspicion that one of Deep Blue's moves had a human touch. It was never proven or disproven, but I trust Kasparov's deep intuition amplified by Kasparov requesting access to Deep Blue’s logs, and IBM refusing to share them in full. For more discussions see [1][2][3].
[1] https://chess.stackexchange.com/questions/9959/did-deep-blue...
[2] https://nautil.us/why-the-chess-computer-deep-blue-played-li...
[3] https://en.chessbase.com/post/deep-blue-s-cheating-move
kinda wild that an llm cant translate to lean?
Both are true. One spent $400 in compute and the other one spent a lot more.
Exactly. And presumably had a more sophisticated harness around the model, longer reasoning chains, best of N, self judging, etc
OpenAI achieved Gold on an unreleased model. GPT-5. Read the tweets and they explain what they did.
Actually, I did it a year ago but I just don't want to release my model.
Where should I address the billion dollar check?
My buddy did it 5 years ago. You wouldn’t know him, he lives in Canada.
my model goes to a different school
The dog ate mine. And the solution didn't fit in the margin, anyway.
OpenAI explicitly said it’s not GPT-5 but another experimental research model https://x.com/alexwei_/status/1946477756738629827?s=46
Thanks. I parsed that wrong. In either case not the same thing Math Arena used.
"You know that really hard test thing that most humans on the planet can't do, or even understand, yeah, LLMs kind of suck at it too"
Meanwhile Noam "well aschtually..."
I love how people are still betting against AI, its hilarious. Please write more 2000-esk "The internet is a fad" articles
Its quite reasonable. We have yet to meet anything more intelligent than humans so why do we think we can create something more intelligent than us when we don't fully understand the complexities how we work?
AI still has a long way to go, though it has proven to be a useful tool at this point.
who said anything about creating something more intelligent than us, these articles have the air of "why are we wasting our time on this stuff", people like gary marcus link them, meanwhile they get better week over week