But to many commentors he's a now gatekeeping AI-hating Luddite clinging to a dying profession out of bitterness and envy because his position is more nuanced than "throw AI at everything and turn off your brain".
Lee Sedol said in an interview that "losing to AI, in a sense, meant my entire world was collapsing. ... I could no longer enjoy the game. So I retired", and I think there will be folks in the mathematical community who would feel the same when the solutions pages to hard problems are suddenly available.
But on the other hand, people learned a lot from chess engines. After decades of chess computers beating humans, there was still a renewed interest in watching Leela beat Stockfish, with many people trying to understand the strategy Leela used.
If your happiness comes from grinding on a problem and making progress, the prospect of having to dig through a corpus of AI-generated proofs might be hard to swallow. But if you're willing to do that, you will still find beautiful things that only so many people can truly appreciate.
There is a real, undeniable possibility of AI becoming better at mathematics in the same way that it became better at chess and Go, and in such a scenario, one may expect the community's response to be comparable.
Tech companies are as much the topic of this post as AI, I think that's the immediacy.
I’ve watched quite a lot of YouTube videos where two machines compete, so you may not be completely right here
Top chess engine championship is pretty fun to watch.
Stockfish does not steal research or scoop researchers.
The concentration of computing resources and capital should be examined by the math community.
Carlsen is bored by studying engine lines.
The popularity is boosted by YouTubers because chess is very suitable for somewhat higher class content.
I'm not sure we'd want that world for math. Positions will be cut just like archaeologist positions are cut now.
I'm not sure what the equivalent would look like in the math field, but it probably involves a lot of mathematicians losing their jobs and the quality of human-produced math decreasing overall.
The quality of the math in general would be fine, since in this scenario cpus will keep producing it. The quality of cpu-cpu chess games is quite high, beyond human understanding in many cases.
Chess is a weird example because it doesn't really have any utility beyond itself. Even pure math sometimes ends up having use in the strangest places. Although if no one understands the frontier math (because no one is getting paid to), I'm not sure it even matters what the quality of the cpu math is?
It's a bit like a tree falling in a forest. If an LLM proves a theorem but no one understands it, did it make a sound?
Also I'd argue that chess's utility is ultimately the same as pure math, particularly in esoteric fields. These things are highly improbable to ever lead to any sort of real world breakthrough. The main benefit is an outlet for human logic, creativity, and exploration - which significant self improvement possible along the journey for players.
Presumably AI will be connect the dots to the applications. As the declaration says, this isn't just about math. Human understanding is losing economic value. You can understand stuff on your own time, I guess.
The standard justification for pure math to holders of purse-strings is something like "it might lead to a useful application down the road, like crypto, who knows". That looks pretty inefficient now. We have to entertain the possibility that AI can develop the math needed for any application we put to it. Eg if number theory didn't exist, we could have asked AI for a way to transit messages securely and it would maybe come up with fermats little theorem as part of its solution or maybe come up with an approach we can't conceive of right now seeing as most of us are constrained to available number theory. Like how in the last year when I give an LLM a programming project I see it doesnt even bother with of the many software libraries I and others have written and just codes up the calls it needs on the fly or finds some other ad hoc solution.
But in future most proofs will be for consumption by other AI models in the pursuit of yet other proofs.
It's kind of surprising so many mathematicians act surprised by this given this was clearly where automated proof assistants would lead. I guess they assumed they'd always be the ones guiding them.
What is the purpose of that?
Its like art being produced for AI to consume. What is gained from that?
And at some point AI will start suggesting - or doing - physical experiments.
But if AI is to recursively self improve understanding and evolving its own foundations, which are clearly mathematical, is essential. There is no need for humans to grasp what is going on in that loop.
Yes.
Friends and I often work on Putnam problems and this series:
The (Almost) Impossible Integrals, Sums, and Series by Cornel Ioan Vălean
There's two points about this I am assuming 1) Mathematics actually has a significant subjectivity to it and is community oriented and not just climbing a never ending list of theorems that exists in the universe 2) A lot of mathematical research work is inside of a subfield and isn't directly motivated by applications. Sometimes it is but e.g. people don't work on obscure theorems about elliptic curves because of a dire need for that but more because the community found it interesting.
Does your "one" only contain humans or does it also contain other AI systems. AI math is not a single monolithic thing, but a distributed one. I see value in sharing proofs even among just AI.
If your goals are understanding the game, self improvement, building thinking skills-- this is the best chess has ever been. It's only if your goal is to beat every opponent you can find that chess is in a bad place.
Even if you don't blunder anything, you'll still find yourself in a worse position without any clue as of what went wrong and why.
Whereas when playing humans, they can usually explain their approach and when they noticed errors in your play.
It's the same for Magnus Carlsen. Even with Queen odds, Stockfish is literally unbeatable for the best players in the world. It's just too strong at evaluating all kinds of random tangent moves (and ensuing positional advantage) which no human player can possibly pay attention due to the time required.
Stockfish vs any human is like Carlsen vs other players by about 3-5 orders of magnitude[0]. It's that stark.
[0] A wild pun appears.
EDIT: To avoid having to respond to each responder, fair comments about Queen odds. Maybe I was thinking Rook odds? Also, I kinda lumped Stockfish in with all the other engines, but I realize there are other engines with different properties ofc.
Leela odds networks, on the other hand, are an entirely different beast. I cannot beat Leela queen odds, much less rook or minor piece odds, and even GMs struggle against Leela knight odds.
Without odds though, yeah, Stockfish is just incomprehensibly strong by human standards. All top chess engines are, but Stockfish moreso.
But sometimes, these GMs can flag it, which counts as a win (especially when it's proxied by a cheater). Sometimes they can also explain the idea that cost them the game, so they've learned something maybe.
Whereas us scrubs literally cannot do anything at all for reasons completely beyond our understanding.
Computer moves are typically much more concrete than human moves: a human will play based on pattern matching ("intuition") and can only make explicit calculation of a small fraction of possibilities, after which decisions are guided by guesswork. The computers are unbeatable in practice because they can calculate concretely in seconds what might take an expert human long intensive study to notice, and they don't make the same kinds of oversights humans can make.
But if you stop and explore a particular position for an extended time, and if you have an intermediate level of chess skill, you too can probably often (usually?) figure out why it's doing something. Sometimes understanding the computer's reasons takes searching multiple branches of a tree several unlikely looking moves deep, but the collection of threats the computer was preemptively thwarting, traps it was setting, etc. are comprehensible to humans with enough effort, especially in games between the computer and a human.
The frustrating thing about playing against the computer is that it notices and thwarts every plan you might come up with, before you make up the plan yourself, and it doesn't make (human-apparent) mistakes, so the game ends up feeling hopeless. Nothing you try works on it, and if your idea is even slightly inaccurate it will be exploited.
There was a renaissance during Covid and due to 'The Queen's Gambit' where it gained much more mainstream popularity, but... Chess AI was already far far (like 1000+ Elo) ahead of human players at that point.
The thing is... chess is humans playing (communicating) with humans and that's what keeps it interesting. Check out the view counts of chess AI tourneys vs. human tourneys.
It would be more like Lem's novel where it completely disappears from the human horizon: https://en.wikipedia.org/wiki/Golem_XIV
Perhaps indeed a better understanding of what intelligence really is would allow this sort of Uplift (as in Brin's books).
In the second case all human-level maths would be solved and what lies beyond would be always out of our scope.
This also sounds like a vector for trolling the community with complex putative proofs hiding a known flaw.
Say that AI gives you a Lean proof and says it proves Theorem X. It could just as easily give you the same proof but claim that it proves (not X). How would you know the difference?
Nothing can really be considered proven unless a human expert can read the Lean proof and determine that (X as defined in the Lean proof) corresponds to X. The proof (at least the statement of the theorem) must be intelligible to humans to have value.
It's possible people will just start taking AI at its word. Maybe AI says "Here is a Lean proof of X" and we all just shrug and go "Okay, X is proven." But that's not how it works right now for human mathematicians. Why would we apply that standard for AI?
Edit: Yup. A bug report to Lean was disguised as a "Collatz" proof in a humorous way. Links below.
Dr. Tao said the same thing. Somehow, this letter came through. He wants to conduct Math competitions where participants who don’t have formal credentials can contribute to mathematical research through AI.
Title: Terence Tao - SAIR Competitions and the Future of Experimental Mathematics
https://www.youtube.com/watch?v=rB9YOi3lb7w
and this:
Daniel Litt - Working with LLMs to do high quality math
And embarrassingly they used him for a "coal miners should learn math" moment that just benefits the AI industry.
He has severely reversed course in the past week. Without concrete propositions it remains to be seen how much of the new resistance is for show.
AI is a tool. It speaks languages I don't (Math, Science, Code). I would love to participate in a Math competition without a hint of any formal advanced math training because my experience so far tells me I will do well.
Apparently he has since changed his mind.
He doesn't see value in scrolling through unsolved problems asking an AI to please solve them. In his view, this is a fundamental confusion about what mathematical research is for. Knocking down unsolved problems without developing the community's understanding of them is like prompting Claude to go through a Jira board, write code for all the open tickets, and then close them without merging or deploying the code.
Yet that's exactly how the field works. A new grad student is tasked with finding a suitably difficult problem from a list of unsolved problems. The sweet spot is obscure, so that fewer people are working on it, but not too obscure that no one knows about it. It works the same way in theoretical physics and theoretical Comp Sci, and I speak from insider knowledge. The rosy view of mathematicians in the media is largely a product of marketing.
(A) it furthers human knowledge
(B) it gets used in applied sciences, engineering, etc.
Thank You!
Professor Tao.
I don't like nuance here. I think progress is really measured by what humans are able to do and understand, not machines. It is significant if we find problems we struggle to solve. That tells us something. What does it take for humans to solve these problems is related.
The best analogy I can give is if you wanted to climb Mt. Everest you might ask someone for guidance. Would it be better to ask someone who has climbed Mt. Everest or someone who took a helicopter ride up near the top and then went to the peak? This is like the AI versus human gap to me. The helicopter is like using AI to generate a proof. The person who actually climbed Mt. Everest has firsthand knowledge of the experience. Same thing for a difficult proof. The struggle people have is actually valuable here. Likewise, we know people are actually capable of climbing Mt. Everest but if they had only ever rode a helicopter to the top, the knowledge of climbing it would not exist, and surely that is meaningful knowledge given the risks.
So if we rely on AI for proofs I think we lose a sense of what is difficult and why. We lose a sense of what human achievement is. Surely climbing Mt. Everest means more than taking a helicopter up? For students, why bother grinding through all the material of climbing Mt. Everest and then attempting it if the helicopter ride is how things are done now? This would have the affect of destroying knowledge.
(please do not nitpick the analogy because it's the best but perhaps a clumsy way to describe my thoughts)
> From my memory I think they said it took 88 hours to solve a Millenium Problem versus the decades of time humans have put into it.
Keep in mind those ~88 hours were spread across ~10,000 simultaneous agent instances.
So, roughly 880,000 hours of compute.
Assuming a fifty-year career, and forty-hour workweeks, a human mathematician's career is about 100,000 hours of "compute".
I suspect that with six good mathematicians spending their whole careers primarily focused on it, and working together closely, Navier-Stokes might well have fallen already.
The perverse incentives of academia mean this has never occurred.
The perverse incentives of industry mean OpenAI intentionally scooped researchers who were getting close (granted, with AI help).
I'm not trying to dismiss the achievement - if the proof turns out to be solid, it's quite impressive (though much less so if the training data included the recent human breakthrough, which seems pretty plausible).
I'm just pointing out that "88 hours" is a very misleading way of framing this.
> The perverse incentives of academia mean this has never occurred.
This. Mathematicians in their most energetic years are trying to get tenure or land a tenure-track job. They are disincentivized to go all-in on ultra high risk, high-reward problems. The potential downside is just too forbidding. It's much safer to develop a research program in a mainstream field that affords many opportunities for partial progress that can translate to a robust publication record.
Depends on if I want to go by helicopter myself.
> why bother grinding through all the material of climbing Mt. Everest and then attempting it if the helicopter ride is how things are done now?
I think you answered this yourself earlier:
> I think progress is really measured by what humans are able to do and understand
People want to make this progress. Therefore people will "grind Everest" as a mathematical community, and that is maybe not so hugely different from a lot of previous mathematical work.
There's still ample room for creativity: simplifying, generalizing, asking new questions humans are interested in, ...
Building a machine that solves Millennium problems is pretty cool too. You wouldn't know it from reading these stories, though.
I think a better comparison is: mathematics just becomes like mining bitcoins.
In reality, the value of Bitcoin is determined by humans (even if indirectly, not by planning), and I think the OP’s point may have been that maths proofs can be regarded similarly. No intrinsic value, just what humans find in it.
Why would research be closed in one direction? Even if AI or human says "Tried that, didn't work" or whatever, someone (or something I suppose) might very well retry it in the future, if nothing else to reproduce it didn't work, in theory at least.
The current wave of AI slop mathematics might end up driving the next generation of mathematicians away from the subject for the same reason that Mochizuki would have convinced me to quit if his proof had been accepted by the community. Luckily, my professors had the taste to immediately recognize that it was garbage.
Is there an established term for the idea of "DoS"? I've taken to calling it slop fatigue.
Mochizuki was still one human and it required legions of other humans to unpack and untangle to confirm that it didn't lead to anywhere in particular.
AI is now capable of constructions so complex that no human or human team can unpack. And its ability to increase that complexity is growing while our human ability is stagnant.
meta-AI analysis cannot help. We (software professionals who use AI regularly) already know that if you run into a situation where a Fable/Astra-generated analysis reaches the limits of our comprehension/complexity due to their subjectivity, throwing more AI at the problem doesn't always converge.
There are many reasons to feel optimistic about AI, and ultimately its general ability to help science and mathematics.
I see no reason to feel optimistic about the future of mathematics and AI based on the current path of frontier labs, unless the misalignment Tao is writing about can be reconciled.
How can we possibly know this when we haven't even seriously started on the endeavor of actively reverse engineering these AI-generated proofs? That's a proper job for human mathematicians, because the AIs themselves are demonstrably clueless about what steps in a proof are genuinely interesting and load-bearing from a human POV. This is evidence of a limitation in AIs' capabilities, not of any kind of misaligned behavior. The fact that Tao actually uses that term in his complaint is deeply disappointing.
Algorithmic verification is a very unsatisfying answer to the problem (e.g., surely it's not just dumb luck that every single case happen to have this exact property), but that's an entirely different issue than saying that no one follows logic of the proof method itself.
Everyone knows that debugging is twice as hard as writing a program in the first place. So if you're as clever as you can be when you write it, how will you ever debug it?
It's prescient.
Could AI write programs that humans can’t understand or debug? Probably, but that’s not what Kernighan was describing.
Can you give an example of this?
It'll basically become slop fatigue if OpenAI starts dumping out proofs faster than the community can keep up, and some turn out to be wrong, never formalize it, don't stay to support it, etc.
(Well that's my hopeful, optimistic take, anyway.)
That said, that’s probably just because of the drama miring their most recent one. After 2 I don’t see why they’d bother anymore.
P/NP and the Riemann Hypothesis are part of the milllenium problems. They will 100% keep trying to crack those regardless.
Imagine a world where these most complex mathematical problems are not accessible to a few hundred people, but a few hundred thousands people. ...Those original few hundred gifted mathematicians would have an even more prominent role, and their names and achievements would be known by orders of magnitude more people that they are now.
Based on current reward models, the frontier AI labs will burn down mathematics as an impressive display of capabilities and in doing so, will make it impossible for people that get paid to do mathematics to stay employed.
If your job is literally to publish papers, and OpenAI and Anthropic decide that making an infinite-paper-printing machine is the best thing to show how effective their tech is, then as a demo, they destroy that industry.
I'm not a fan of knocking down things that work, however I also find it hard to be against death of the gatekeeping old guard of any industry.
I think math is just gonna have to suck it up like every other industry now. Math productivity is longer out of reach of the average grad student. Like every other industry they are no longer untouchable and are gonna have to adjust to the new way of things or market forces will do what they always do which is refuse to fund ineffectiveness.
I've had to accept that tech/IT will never be the same. Just how it is. You can thrash against it all you want.