nybsjytm

I called it the easiest part of his papers, not easy. Either way, it's actually not relevant to the proof. For example, I believe that Morgan and Tian's 500 page exposition of the proof doesn't mention it even once.

Moreover I'd strongly dispute that there was any particular point where it was clear that the problem was "near a solution." The W functional is an example where experts could very quickly verify that Perelman had made at least one major new discovery about the Ricci flow. But the proof of the Poincaré conjecture was on another order of complexity and required different (more complicated) new results about Ricci flow.

> brilliantly realised

Can you say more about this? Nothing about this approach seems very amazing to me. Construct an approximate solution by some numerical method (in this case neural networks), prove that a solution which is close enough to satisfying the equation can be perturbed to an exact solution. Does the second half use some nonstandard method?

> Folks knew the problem was near a solution once the monotonicity proof of the W functional came out.

This isn't true, it was a major accomplishment but by far the easiest part of Perelman's papers and not actually even part of the proof of the Poincaré conjecture.

> PINNs are different in concept, yes, but clearly no less important

If anything I think they're more important! Whether or not it works out for Navier-Stokes, this kind of thing is an extremely plausible avenue of approach and could yield interesting singularities for other major equations. I am however extremely concerned about public understanding. I know you are well aware that this is worlds away from the speculative technologies like 'mathematical superintelligence' but, if it works out, it'll be like a nuclear bomb of misinformation about AI and math.

> I know they are so close to a computationally-assisted proof of counterexample that it is virtually inevitable at this point.

That's a strong claim. Is it based on more than the linked work on some model problems from fluid mechanics?

I will say that I dread the discourse if it works out, since I don't believe enough people will understand that using a PINN to get new solutions of differential equations has substantially no similarity to asking ChatGPT (or AlphaProof etc) for a proof of a conjecture. And there'll be a lot of people trying to hide the difference.

Not sure what this has to do with my post.

Any mathematicians who have actually called it "new interesting mathematics", or just an OpenAI employee?

The paper in question is an arxiv preprint whose first author seems to be an undergraduate. The theorem in it which GPT improves upon is perfectly nice, there are thousands of mathematicians who could have proved it had they been inclined to. AI has already solved much harder math problems than this.

There's a related section about 'mathiness' in section 3.3 of the article "Troubling Trends in Machine Learning Scholarship" https://arxiv.org/abs/1807.03341. I would say the situation has only gotten worse since that paper was written (2018).

However the discussion there is more about math which is unnecessary to a paper, not so much about the problem of math which is unintelligible or, if intelligible, then incorrect. I don't have other papers off the top of my head, although by now it's my default expectation when I see a math-centric AI paper. If you have any such papers in mind, I could tell you my thoughts on it.

Much physicist math can't be made rigorous so easily! Which isn't to say that much of it doesn't still have great value.

However the math in AI papers is indeed different. For example, Kingma and Ba's paper self-presents as having a theorem with a rigorous proof via a couple of lemmas proved by a chain of inequalities. The key thing is that the mathematical details are purportedly all present, but are just wrong.

This isn't at all like what you see in physics papers, which might just openly lack detail, or might use mathematical objects whose existence or definition remain conjectural. There can be some legitimate problems with that, but at least in the best cases it can be very visionary. (Mirror symmetry is a standard example.) By contrast I'm not sure what 'spirit' is even possible in a detailed couple-page 'proof' that its authors probably don't even fully understand. In most cases, the 'theorem' isn't remotely interesting enough as pure mathematics and is also not of any serious relevance to the empirical problem at hand. It just adds an impressive-looking section to the paper.

I do think it's possible that in the future there will be very interesting pure mathematics inspired by AI. But it hasn't been found yet, and I'm very certain it won't come from reconsidering these kinds of badly-written theorems and proofs.

> Checking the correctness of proofs is a much easier problem than coming up with the proof in the first place.

Just so this isn't misunderstood, not so much cutting-edge math is presently possible to code in lean. The famous exceptions (such as the results by Clausen-Scholze and Gowers-Green-Manners-Tao) have special characteristics which make them much more ground-level and easier to code in lean.

What's true is that it's very easy to check if a lean-coded proof is correct. But it's hard and time-consuming to formulate most math as lean code. It's something many AI research groups are working on.

> Many AI researchers are mathematicians. Any theoretical AI research paper will typically be filled with eye-wateringly dense math. AI dissolves into math the closer you inspect it. It's math all the way down.

There is a major caveat here. Most 'serious math' in AI papers is wrong and/or irrelevant!

It's even the case for famous papers. Each lemma in Kingma and Ba's ADAM optimization paper is wrong, the geometry in McInnes and Healy's UMAP paper is mostly gibberish, etc...

I think it's pretty clear that AI researchers (albeit surely with some exceptions) just don't know how to construct or evaluate a mathematical argument. Moreover the AI community (at large, again surely with individual exceptions) seems to just have pretty much no interest in promoting high intellectual standards.

I consider it unconfirmed until it happens! No idea where I saw it but it was probably on twitter.

This takes for granted a formal setting, which is what I'm questioning in any of these 'real world' contexts.

Hmmm I think even in something very nominally nearby like theoretical physics, there's very little that's similar to theorem proving. I don't see how AlphaProof could be a stepping stone to anything like what you're describing.

Generally, I think many people who haven't studied mathematics don't realize how huge the gulf is between "being logical/reasonable" and applying mathematical logic as in a complicated proof. Neither is really of any help for the other. I think this is actually the orthodox position among mathematicians; it's mostly people who might have taken an undergraduate math class or two who might think of one as a gateway to the other. (However there are certainly some basic commonalities between the two. For example, the converse error is important to understand in both.)

You don't even need AI to regurgitate Perelman's papers, you can do that in three lines of python.

What I meant is that there's no AI you can ask to explain the details of Perelman's proof. For example, if there's a lemma or a delicate point in a proof that you don't understand, you can't ask an AI to clarify it.

Very plausible, but that would also be noteworthy. As I've mentioned in some other comments here, (as far as I know) we outside of DeepMind don't know anything about the computing power required to run alphaproof, and the tradeoff between computing power required and the complexity of problems it can address is really key to understanding how useful it might be.

If this were the case, I don't see why we'd need to wait for an AI company to make a breakthrough in math research. The key issue instead is how to encode 'real-life' statements in a formal language - which to me seems like a ludicrous problem, just complete magical thinking.

For example, how might an arbitrary statement like "Scholars believe that professional competence of a teacher is a prerequisite for improving the quality of the educational process in preschools" be put in a lean-like language? What about "The theoretical basis of the October Revolution lay in a development of Marxism, but this development occurred through three successive rounds of theoretical debate"?

Or have I totally misunderstood what people mean when they say that developments in automatic theorem proving will solve LLM's hallucination problem?

The quality of AI algorithms is not based on formal mathematics at all. (For example, I'm unaware of even one theorem relevant to going from GPT-1 to GPT-4.) Possibly in the future it'll be otherwise though.

No, nobody has proved it.

Side point, there is no existing AI which can prove - for example - the Poincaré conjecture, even though that has already been proved. The details of the proof are far too dense for any present chatbot like ChatGPT to handle, and nothing like AlphaProof is able either since the scope of the proof is well out of the reach of Lean or any other formal theorem proving environment.

As a mathematician, of course I agree. But in a sentence like:

> A speedup in the movement of the maths frontier would be worth many power stations

who is it 'worth' it to? And to what end? I can say with some confidence that many (likely most, albeit certainly not all) mathematicians do not want data centers and power stations to guzzle energy and do their math for them. It's largely a vision imposed from without by Silicon Valley and Google research teams. What do they want it for and why is it (at least for now) "worth" it to them?

Personally, I don't believe for a second that they want it for the good of the mathematical community. Of course, a few of their individual researchers might have their own personal and altruistic motivations; however I don't think this is so relevant.

Being logical in social life is pretty much completely different from being logical in a mathematical argument, especially in a formal theorem proving environment. (Just try to write any kind of cultural proposition in a formal language!)

It's worth emphasizing that it's been possible for years to use an automatic theorem prover to prove novel results. The whole problem is to get novel interesting results.

> A better question is what can happen when everybody has access to above average reasoning. Our society is structured around avoiding confronting people with difficult questions, except when they are intended to get the answer wrong.

What does this have to do with a hypothetical automatic theorem prover?

For some time a 'superhuman math AI' could be useful for company advertising and getting the attention of VCs. But eventually it would be pretty clear that innovative math research, with vanishingly few exceptions, isn't very useful for making revenue. (I am a mathematician and this is meant with nothing but respect for math research.)

I agree that the result is important regardless. But the tradeoff of computing time/cost with problem complexity is hugely important to think about. Finding a proof in a formal language is trivially solvable in theory since you just have to search through possible proofs until you find one ending with the desired statement. The whole practical question is how much time it takes.

Three days per problem is, by many standards, a 'reasonable' amount of time. However there are still unanswered questions, notably that 'three days' is not really meaningful in and of itself. How parallelized was the computation; what was the hardware capacity? And how optimized is AlphaProof for IMO-type problems (problems which, among other things, all have short solutions using elementary tools)? These are standard kinds of critical questions to ask.

There are some AI guys like Christian Szegedy who predict that AI will be a "superhuman mathematician," solving problems like the Riemann hypothesis, by the end of 2026. I don't take it very seriously, but that kind of prognostication is definitely out there.

This has to come with an asterisk, which is that participants had approximately 90 minutes to work on each problem while AlphaProof computed for three days for each of the ones it solved. Looking at this problem specifically, I think that many participants could have solved P6 without the time limit.

(I think you should be very skeptical of anyone who hypes AlphaProof without mentioning this - which is not to suggest that there's nothing there to hype)

Why have they still not released a paper aside from a press release? I have to admit I still don't know how auspicious it is that running google hardware for three days apiece was able to find half-page long solutions, given that the promise has always been to solve the Riemann hypothesis with the click of a button. But of course I do recognize that it's a big achievement relative to previous work in automatic theorem proving.

Hinton was never in the running for a Fields medal since he never made a single contribution to the field of mathematics. His work is about empirical discoveries in CS.

An incomprehensibly silly publicity stunt!