Thank you for the very interesting post. But I find it hard to follow your conclusions. Sure the AI labs have managed to solve many open problems but also failed to resolve many more despite the ungodly amount of material and human resources they threw at them. Should mathematicians just give up on those problems either in the hope/anticipation that some future iteration of these models will be able to resolve them or because the failure of AI supposedly shows that these are beyond reach for humans as well? And of course this assumes that AI labs will in the foreseeable future continue to invest billions into solving pure math problems. The same goes for the claim of limitless AI driven progress - arguably many of the results produced so far have been last mile results along the lines of established research efforts and AI has yet to prove that it can produce truly original contributions/insights to mathematics comparable to the milestones/breakthroughs of the past that not only resolved existing problems/conjectures but created entirely new subfields of mathematics.
>Should mathematicians just give up on those problems ...
The models are jagged even within the various field of maths . There should be problems that the average mathematician could solve that the models can't.
Tho probably a temporary state of affairs.
>And of course this assumes that AI labs will in the foreseeable future continue to invest billions into solving pure math problems
Noam brown has mentioned that it has become harder to find hard problems that can be benchmarked and trained on but still within reach of current models.
Solving pure math problems is an easy way to find out so I think they will continue until the current collection of problems runs out. They could create their own problems if that happens
There could be a huge cliff in that unsolved problems could be too difficult for the models stalling progress but less likely imo
>AI has yet to prove that it can produce truly original contributions/insights to mathematics
Creating new fields of math is not the same as mere problem solving but I think it is a limitation of current models than a fundamental one that will not be cleared
Couple of years ago the models were unable to solve basic elementary questions so I reserve some humility to not think that future models won't be able to do X.
i think we are still in the centaur stage even now, and will be unless there is a true intelligence explosion.
as an ordinary doofus, if I just follow the moves of a chess engine i can beat anyone. i could do this without even understanding the rules of chess.
whereas with math, it requires deep mathematical training to even formulate a question, and to make any sense of the result or do anything useful with it at all. it is hard to imagine a world where a layperson, or even a mathematically-trained engineer, could be anywhere near as effective as a real mathematician at using these AI tools to solve meaningful problems.
there would have to be an AI system which is capable not only of solving mathematical problems, but also of educating an ordinary person up to the cutting edge of mathematical knowledge very quickly so that they can understand their own problem properly. this would have large consequences.
if there hadn't been so many famous open problems carefully posed and analyzed in the literature, would OpenAI technical staff have been able to conceive of so many consequential open problems themselves? I don't think so. This key part of the work was done by the mathematical community already.
what will be hard to maintain is the prestige economy where mathematicians invent open problems and then solve them, convincing authority figures that they have accomplished something impressive. but inventing open problems will still be useful, they just won't stay open any more.
i guess it's more humanities-like but this is still very far from a humanities field, mathematicians' actual understanding of mathematics will remain very practically important and desirable.
Do you know whether OpenAI trawled carefully through open math problems, rather than just telling a model to find a difficult problem and work on it? I asked an AI for some unsolved problems on the level of the ones just released. I'm no mathematician, judge for yourself:
1. The smooth 4-dimensional Poincaré conjecture is false: there is an exotic smooth S⁴.
2. The Euler–Mascheroni constant γ is irrational.
3. The lonely runner conjecture holds for any number of runners.
4. The Černý conjecture holds: every synchronizing n-state automaton has a reset word of length at most (n − 1)².
5. The Andrews–Curtis conjecture is false: there is a balanced presentation of the trivial group that cannot be reduced to the trivial presentation by AC moves.
If so, I can use AI to generate the problems to work on and AI to solve them.
In the recent PHP vuln discovery, one company got alerted that there was an upcoming critical announcement about PHP. With that knowledge, they asked an AI to look for bad vulnerabilities in PHP, and it independently discovered the issue.
Probably a professional mathematician can get farther than I can today, but you don't *need* a human to figure out some of the hardest problems. Soon any average Joe can just say "figure out a Nobel Prize problem" and have a decent chance at getting a good answer.
There is no mathematics without humans. It's a practice not a result. Working on problems is the major part of the process of mathematics and discovering new ideas and creating new mathematicians with new ideas and problems. It's quite obvious this will just kill the field altogether, and set us back centuries. It will not create any new math.
I think this analysis stops at a slightly strange place, and I do really respect it for its depth of experience. But if the coming age is an age where all cognitive fields become humanities (and notably, if you assume most of this progress does not change anything about the currently-fundamental capability and generality constraints in robotics, which I don’t agree is true), who will pay for this? How many people will have this as their livelihood?
I don’t fully agree with the characterization of the humanities as fields where no advancement of the frontier is expected. But even if so, it seems unclear that we’d be able to employ our former mathematicians, our former physicists, our former engineers to an equal number of roles as interpreters and scribes.
Don’t a bunch of people end up unemployed in the transition, even assuming there is somewhere to transition to? And doesn’t that have large societal impacts, if it happens?
Thank you for the very interesting post. But I find it hard to follow your conclusions. Sure the AI labs have managed to solve many open problems but also failed to resolve many more despite the ungodly amount of material and human resources they threw at them. Should mathematicians just give up on those problems either in the hope/anticipation that some future iteration of these models will be able to resolve them or because the failure of AI supposedly shows that these are beyond reach for humans as well? And of course this assumes that AI labs will in the foreseeable future continue to invest billions into solving pure math problems. The same goes for the claim of limitless AI driven progress - arguably many of the results produced so far have been last mile results along the lines of established research efforts and AI has yet to prove that it can produce truly original contributions/insights to mathematics comparable to the milestones/breakthroughs of the past that not only resolved existing problems/conjectures but created entirely new subfields of mathematics.
Disclaimer : Not a mathematician
>Should mathematicians just give up on those problems ...
The models are jagged even within the various field of maths . There should be problems that the average mathematician could solve that the models can't.
Tho probably a temporary state of affairs.
>And of course this assumes that AI labs will in the foreseeable future continue to invest billions into solving pure math problems
Noam brown has mentioned that it has become harder to find hard problems that can be benchmarked and trained on but still within reach of current models.
Solving pure math problems is an easy way to find out so I think they will continue until the current collection of problems runs out. They could create their own problems if that happens
There could be a huge cliff in that unsolved problems could be too difficult for the models stalling progress but less likely imo
>AI has yet to prove that it can produce truly original contributions/insights to mathematics
Creating new fields of math is not the same as mere problem solving but I think it is a limitation of current models than a fundamental one that will not be cleared
Couple of years ago the models were unable to solve basic elementary questions so I reserve some humility to not think that future models won't be able to do X.
Hu, how do we know that the proofs were obtained mostly through non-formalized reasoning?
I'm not sure as well I'm convinced of this point in particular.
I thought you might be interested in my post that is riffing off of this: https://carlvphillips.substack.com/p/math-is-the-ultimate-liberal-art
i think we are still in the centaur stage even now, and will be unless there is a true intelligence explosion.
as an ordinary doofus, if I just follow the moves of a chess engine i can beat anyone. i could do this without even understanding the rules of chess.
whereas with math, it requires deep mathematical training to even formulate a question, and to make any sense of the result or do anything useful with it at all. it is hard to imagine a world where a layperson, or even a mathematically-trained engineer, could be anywhere near as effective as a real mathematician at using these AI tools to solve meaningful problems.
there would have to be an AI system which is capable not only of solving mathematical problems, but also of educating an ordinary person up to the cutting edge of mathematical knowledge very quickly so that they can understand their own problem properly. this would have large consequences.
if there hadn't been so many famous open problems carefully posed and analyzed in the literature, would OpenAI technical staff have been able to conceive of so many consequential open problems themselves? I don't think so. This key part of the work was done by the mathematical community already.
what will be hard to maintain is the prestige economy where mathematicians invent open problems and then solve them, convincing authority figures that they have accomplished something impressive. but inventing open problems will still be useful, they just won't stay open any more.
i guess it's more humanities-like but this is still very far from a humanities field, mathematicians' actual understanding of mathematics will remain very practically important and desirable.
Do you know whether OpenAI trawled carefully through open math problems, rather than just telling a model to find a difficult problem and work on it? I asked an AI for some unsolved problems on the level of the ones just released. I'm no mathematician, judge for yourself:
1. The smooth 4-dimensional Poincaré conjecture is false: there is an exotic smooth S⁴.
2. The Euler–Mascheroni constant γ is irrational.
3. The lonely runner conjecture holds for any number of runners.
4. The Černý conjecture holds: every synchronizing n-state automaton has a reset word of length at most (n − 1)².
5. The Andrews–Curtis conjecture is false: there is a balanced presentation of the trivial group that cannot be reduced to the trivial presentation by AC moves.
If so, I can use AI to generate the problems to work on and AI to solve them.
In the recent PHP vuln discovery, one company got alerted that there was an upcoming critical announcement about PHP. With that knowledge, they asked an AI to look for bad vulnerabilities in PHP, and it independently discovered the issue.
Probably a professional mathematician can get farther than I can today, but you don't *need* a human to figure out some of the hardest problems. Soon any average Joe can just say "figure out a Nobel Prize problem" and have a decent chance at getting a good answer.
There is no mathematics without humans. It's a practice not a result. Working on problems is the major part of the process of mathematics and discovering new ideas and creating new mathematicians with new ideas and problems. It's quite obvious this will just kill the field altogether, and set us back centuries. It will not create any new math.
I think this analysis stops at a slightly strange place, and I do really respect it for its depth of experience. But if the coming age is an age where all cognitive fields become humanities (and notably, if you assume most of this progress does not change anything about the currently-fundamental capability and generality constraints in robotics, which I don’t agree is true), who will pay for this? How many people will have this as their livelihood?
I don’t fully agree with the characterization of the humanities as fields where no advancement of the frontier is expected. But even if so, it seems unclear that we’d be able to employ our former mathematicians, our former physicists, our former engineers to an equal number of roles as interpreters and scribes.
Don’t a bunch of people end up unemployed in the transition, even assuming there is somewhere to transition to? And doesn’t that have large societal impacts, if it happens?
I wonder if the near-future field to be is in applied maths - taking the AI-generated math and creatively look for application to existing problems.