There are 2 strong points buried between the snark: that the AI math advisory group, which OpenAI announced working with just 2 1/2 weeks ago, is against using internal models to solve open problems (per their general guidelines[0]); and that this mode of dumping results before having a friendly formulation doesn't center human understanding.
IMO OpenAI's way of publishing also goes against other parts of the general guidelines. Also, I doubt they'll follow point 2B3:
> For each result released, the AI lab should make public the name of the model, the prompts used, a (summarized) chain of thought, the time taken, and the estimated cost of computation. Releasing additional material that sheds light on the scientific process that produced the results, such as the initial LLM outputs before they were cleaned up (as recommended in (1) above), is strongly encouraged.
I'm having a really hard time seeing how this letter could conceivably be perceived as persuasive. In fact, this letter makes it clear that the people behind it are seeking to maintain their previously entrenched power, prestige, and control over a field of human knowledge. And they are going about it in a futile and debasing way.
As I understand it, this is one of the groups OpenAI claimed to be consulting with to improve their process after the Navier Stokes fumble. I think OpenAI might have even said something like “in accordance with this group’s guidance” in the big dump? So this post should be read as distancing themselves from OpenAI and denying any legitimacy OpenAI tried to derive from their association.
Edit: on researching more, I don’t think this is the case; I must have confused them with some other group.
You're not talking about this, but I actually said the same thing about the state of software engineering a few weeks ago. So many fields are facing an existential crisis.
they could have definitely written it better and put more effort into it. As it is, they have real gripes with how openai is doing things and i think they should have concretely listed their objections. Either way, if OpenAI dumps a gazillion proofs, who actually is going to read them, understand them, assess their correctness etc... ? If no one is doing that, is that problem really "solved" ? So OpenAI just dumps proofs and calls it a day, or this trillion dollar company can spend some effort engaging with mathematicians but really OpenAI doesn't really care about Math, they want to prove that their frontier intelligence can just do it all before they IPO. You say the authors are trying to maintain their power, but what is OpenAI doing? Certainly, you are naive if you think they are doing this stuff for the benefit of humanity or a goal of advancing mathematical knowledge.
Who knows, If compute keeps dumping proofs for problems but no one knows their correctness because mathematicians stop investing their time in understanding these proofs, then you can't even really use the results or you can use it in the same sense you use stuff you don't really understand: You trust the author. With humans, there is a certain level of trust in people and institutions - with AI, we aren't really there yet.
For thousands of years mathematics had this aura of being at the pinnacle of human thought and was one of the things that separated us from other species. People like Euler, Gauss, Euclid, Hilbert, Erdos had this aura about that them that they were on the noblest of pursuits - trying to solve problems that had no immediate practical use, and largely needed focused thinking over extended periods of time with the occasional spark of creativity.
Then comes along AI, a mass extinction event to the entire field. It turns out LLMs can think for longer without stopping, have better recall and working memory. What we're seeing now is a 5000-year old set of institutions facing an existential threat because they no longer require humans to be the sole source of labor.
I find the closest analogy to be automobiles displacing the horse-carriage industry in the early 1900s. Horses couldn't protest and rapidly went out of a job, but horse-carriage owners and that industry certainly did.
Jokes aside, and regardless of how anyone perceives their field or industry, I feel pretty much every thought-based industry is up for an existential threat to some degree. Given a bit of time, manual labour fields are looking a bit shaky too.
The real question is not "are we impacted?" but rather "to what extent, and how do we integrate it?"
The problem to be solved was not to spew out maximum lines of text containing some greek letters per hour.
It was to write down iterative progress in conceptualizing, modelling and consistently explaining difficult questions - in a way that advances human understanding. If nobody bothers to read through 50 pages of gibberish that may or may not be containing some insight, then the total problem solving remains zero.
> a particular vision of progress that is informed by history
I imagine there has, for a very long time, not been a need to specifically write down what that field-specific vision is and how exactly history teaches us that there really only is one scientific method. Some future LLM will surely be able to explain this in their place, so mathematicians can stick to their primary fields instead of having to become experts in teaching the theory of teaching.
I'm not a mathematician but an engineer, math adjacent.
I recall once saying something to a colleague even more cringe-worthy than this. Fear of one's own position and stature can make us (speaking for myself) say things we later regret.
I am a former mathematician now working a regular job (not affiliated with OpenAI). Just speaking for myself, I find this statement deeply embarrassing and at odds with the spirit of open inquiry for which the community of mathematics has (nearly) always been a standard bearer.
Isn't the big problem here that volunter work is being exploited by billion dollar companies? Of course it's bad and there should be a solution where openai pays people to verify the information they put out in the public before it ever reaches independent reviewers.
Imagine a world in which problems are mined, proofs are generated, yet no one exists to understand or verify the correctness of these proofs. In software parlance, who is going to review all the vibe coded bs that come your way if you maintain an open source library - is the open source maintainer standing in the way of 'progress'? Really, this whole scheme only works if AI were truly autonomous beings, i.e they have a system where they can verify their proofs, use it to understand which problems they need to tackle next, build understanding which is used by other AI to generate tools to solve physics problems etc... but the way we are approaching things does not work if humans are heavily in the loop.
Well it depends on how humans organize society and what these autonomous AI agents actually become - In an ideal world, i wouldn't mind spending my time riding my bike, reading or doing things i intrinsically value and receiving a check every month and let robots grow crops, administrate our government, cure cancer etc... Of course, the rich humans can decide to have private armies of robots and just enslave the rest of the population or robots can develop awareness and they can decide that they are benevolent dictators and humans must accept whatever they propose or they can decide humans are not worth keeping alive. There are many ways this can play out and no one knows what the future holds.
Software engineers somehow got used to slop grenades, 5 figure PRs at full throttle and shift-left on every issue, but not all professions are taking it lying down, it seems.
700 potential papers dumped is not equal to 700 PRs with a LGTM.
Of course the dinosaurs chose to bury their heads in the sand, and asked that everyone else does too. They've no place in the new world. They could've instead come up with good feedback to improve the paper quality, but they didn't.
„Mathematicians did not ask for this work to be done.”
Since when science works like this?
One point I might slightly agree with - refusal to acknowledge work done with nonpublic models. Otoh in other fields and in history it’s been common to do science with resources unavailable to common people.
Nobody needs permission to solve an open problem. That's the whole point of publishing them. The question for any result is whether it's correct, not who produced it or whether anyone requested it.
Bringing up OpenAI's lawsuits is irrelevant to whether these proofs hold. And calling a release that includes Lean formalizations a "demonstration of power" gets it backwards. Machine checkable proofs are the least "trust me" form of mathematics there is.
There are fair criticisms here. Not every result is formalized, the model can't be reproduced by outsiders, and the massive dump strains review capacity. Those are reasons to demand full formalization, open access to the methods, and help funding human review. They aren't reasons to dismiss correct mathematics or to tell people to stop working on hard problems.
A depressing statement that does nothing to increase people's support of working Mathematicians. It feels like a low-effort political hit as opposed to the balanced writeups we've been reading from top Mathematicians like Terence Tao.
He may accept this point of view, but I've never seen something like this coming from his mouth. Even then, while I'm a fan of Tao, it doesn't mean that I have to agree with everything he does.
There are 2 strong points buried between the snark: that the AI math advisory group, which OpenAI announced working with just 2 1/2 weeks ago, is against using internal models to solve open problems (per their general guidelines[0]); and that this mode of dumping results before having a friendly formulation doesn't center human understanding.
IMO OpenAI's way of publishing also goes against other parts of the general guidelines. Also, I doubt they'll follow point 2B3:
> For each result released, the AI lab should make public the name of the model, the prompts used, a (summarized) chain of thought, the time taken, and the estimated cost of computation. Releasing additional material that sheds light on the scientific process that produced the results, such as the initial LLM outputs before they were cleaned up (as recommended in (1) above), is strongly encouraged.
[0] https://agmai.org/general-sep29/
I'm having a really hard time seeing how this letter could conceivably be perceived as persuasive. In fact, this letter makes it clear that the people behind it are seeking to maintain their previously entrenched power, prestige, and control over a field of human knowledge. And they are going about it in a futile and debasing way.
As I understand it, this is one of the groups OpenAI claimed to be consulting with to improve their process after the Navier Stokes fumble. I think OpenAI might have even said something like “in accordance with this group’s guidance” in the big dump? So this post should be read as distancing themselves from OpenAI and denying any legitimacy OpenAI tried to derive from their association.
Edit: on researching more, I don’t think this is the case; I must have confused them with some other group.
You're not talking about this, but I actually said the same thing about the state of software engineering a few weeks ago. So many fields are facing an existential crisis.
they could have definitely written it better and put more effort into it. As it is, they have real gripes with how openai is doing things and i think they should have concretely listed their objections. Either way, if OpenAI dumps a gazillion proofs, who actually is going to read them, understand them, assess their correctness etc... ? If no one is doing that, is that problem really "solved" ? So OpenAI just dumps proofs and calls it a day, or this trillion dollar company can spend some effort engaging with mathematicians but really OpenAI doesn't really care about Math, they want to prove that their frontier intelligence can just do it all before they IPO. You say the authors are trying to maintain their power, but what is OpenAI doing? Certainly, you are naive if you think they are doing this stuff for the benefit of humanity or a goal of advancing mathematical knowledge.
>Certainly, you are naive if you think they are doing this stuff for the benefit of humanity or a goal of advancing mathematical knowledge.
Certainly you are cynical if you think it won't benefit humanity or advance mathematical knowledge.
Who knows, If compute keeps dumping proofs for problems but no one knows their correctness because mathematicians stop investing their time in understanding these proofs, then you can't even really use the results or you can use it in the same sense you use stuff you don't really understand: You trust the author. With humans, there is a certain level of trust in people and institutions - with AI, we aren't really there yet.
Certainly, you are feigning ignorance about them not caring about this but for IPO hyping.
I kind of see their point.
For thousands of years mathematics had this aura of being at the pinnacle of human thought and was one of the things that separated us from other species. People like Euler, Gauss, Euclid, Hilbert, Erdos had this aura about that them that they were on the noblest of pursuits - trying to solve problems that had no immediate practical use, and largely needed focused thinking over extended periods of time with the occasional spark of creativity.
Then comes along AI, a mass extinction event to the entire field. It turns out LLMs can think for longer without stopping, have better recall and working memory. What we're seeing now is a 5000-year old set of institutions facing an existential threat because they no longer require humans to be the sole source of labor.
I find the closest analogy to be automobiles displacing the horse-carriage industry in the early 1900s. Horses couldn't protest and rapidly went out of a job, but horse-carriage owners and that industry certainly did.
https://ghostsofdc.org/2018/07/16/horse-dealers-claim-automo...
>People like Euler
You met the chap?
Jokes aside, and regardless of how anyone perceives their field or industry, I feel pretty much every thought-based industry is up for an existential threat to some degree. Given a bit of time, manual labour fields are looking a bit shaky too.
The real question is not "are we impacted?" but rather "to what extent, and how do we integrate it?"
Are they?
The problem to be solved was not to spew out maximum lines of text containing some greek letters per hour.
It was to write down iterative progress in conceptualizing, modelling and consistently explaining difficult questions - in a way that advances human understanding. If nobody bothers to read through 50 pages of gibberish that may or may not be containing some insight, then the total problem solving remains zero.
"Mathematicians did not ask for this work to be done."
Gatekeeping much?
"Mathematicians have a particular vision of progress that is informed by history and field-specific considerations."
This cannot seriously have been written by anyone mathematically literate, it's just too embarrassing.
> a particular vision of progress that is informed by history
I imagine there has, for a very long time, not been a need to specifically write down what that field-specific vision is and how exactly history teaches us that there really only is one scientific method. Some future LLM will surely be able to explain this in their place, so mathematicians can stick to their primary fields instead of having to become experts in teaching the theory of teaching.
I'm not a mathematician but an engineer, math adjacent.
I recall once saying something to a colleague even more cringe-worthy than this. Fear of one's own position and stature can make us (speaking for myself) say things we later regret.
I am a former mathematician now working a regular job (not affiliated with OpenAI). Just speaking for myself, I find this statement deeply embarrassing and at odds with the spirit of open inquiry for which the community of mathematics has (nearly) always been a standard bearer.
Oi you got a loicense for that maffs?
Isn't the big problem here that volunter work is being exploited by billion dollar companies? Of course it's bad and there should be a solution where openai pays people to verify the information they put out in the public before it ever reaches independent reviewers.
Doesn’t seem like they are taking it well given this kind of response.
Sounds like sore losers. They got beat at their own game by a computer, and they're mad about it.
It's just sour grapes.
Imagine a world in which problems are mined, proofs are generated, yet no one exists to understand or verify the correctness of these proofs. In software parlance, who is going to review all the vibe coded bs that come your way if you maintain an open source library - is the open source maintainer standing in the way of 'progress'? Really, this whole scheme only works if AI were truly autonomous beings, i.e they have a system where they can verify their proofs, use it to understand which problems they need to tackle next, build understanding which is used by other AI to generate tools to solve physics problems etc... but the way we are approaching things does not work if humans are heavily in the loop.
The strategy seems to be that problem is solved via
A) human displacement B) having to use more AI to check AI outputs
Who wins in both scenario’s? Lol
Well it depends on how humans organize society and what these autonomous AI agents actually become - In an ideal world, i wouldn't mind spending my time riding my bike, reading or doing things i intrinsically value and receiving a check every month and let robots grow crops, administrate our government, cure cancer etc... Of course, the rich humans can decide to have private armies of robots and just enslave the rest of the population or robots can develop awareness and they can decide that they are benevolent dictators and humans must accept whatever they propose or they can decide humans are not worth keeping alive. There are many ways this can play out and no one knows what the future holds.
I fail to see how this organization, AHM, has any relevance.
They do have relevant members, such as Peter Scholze
No reference at the quality of the OpenAI results? Not a mathematician, but I heard demonstrations are written rather poorly.
Note, AHM is "The Association for Human Mathematics"
You don't even need to click on the link to predict the Canute-style tone.
The emotional reaction is understandable, but should not guide the rational response.
Software engineers somehow got used to slop grenades, 5 figure PRs at full throttle and shift-left on every issue, but not all professions are taking it lying down, it seems.
700 potential papers dumped is not equal to 700 PRs with a LGTM.
Some more discussion Terence Tao crosspost: https://news.ycombinator.com/item?id=50000421
Related:
“Math 2.0” will need to value mathematical progress more holistically
https://news.ycombinator.com/item?id=50002008
This is language Paul G warned us about since, I don't know, at least a decade. Literally ticks all the boxes.
Reference?
Of course the dinosaurs chose to bury their heads in the sand, and asked that everyone else does too. They've no place in the new world. They could've instead come up with good feedback to improve the paper quality, but they didn't.
Old man yells at cloud.
„Mathematicians did not ask for this work to be done.”
Since when science works like this?
One point I might slightly agree with - refusal to acknowledge work done with nonpublic models. Otoh in other fields and in history it’s been common to do science with resources unavailable to common people.
Mathematicians need the permission of mathematics, not the other way around. That they didn't ask for it is completely beside the point.
Nobody needs permission to solve an open problem. That's the whole point of publishing them. The question for any result is whether it's correct, not who produced it or whether anyone requested it.
Bringing up OpenAI's lawsuits is irrelevant to whether these proofs hold. And calling a release that includes Lean formalizations a "demonstration of power" gets it backwards. Machine checkable proofs are the least "trust me" form of mathematics there is.
There are fair criticisms here. Not every result is formalized, the model can't be reproduced by outsiders, and the massive dump strains review capacity. Those are reasons to demand full formalization, open access to the methods, and help funding human review. They aren't reasons to dismiss correct mathematics or to tell people to stop working on hard problems.
[dead]
[dead]
this is like an extreme extreme version of how the general regards letter from the UN “denouncing” said “horrible” activity engendered by nation “X”
with no teeth, stuff like this starts to feel a little funny+sad
A depressing statement that does nothing to increase people's support of working Mathematicians. It feels like a low-effort political hit as opposed to the balanced writeups we've been reading from top Mathematicians like Terence Tao.
this was guest posted onto his blog i believe so you may want to hold that thought
He may accept this point of view, but I've never seen something like this coming from his mouth. Even then, while I'm a fan of Tao, it doesn't mean that I have to agree with everything he does.