As we await the latest AI-generated results (whether gurgling through the arXiv, appearing on 4chan, or coming straight from the horse’s mouth), I thought I would put on record some interesting open problems from my own field. I don’t mention these in any way to “move the goalposts”, but rather to be a little more explicit about what makes them interesting. In particular, each of these problems—which I consider significant—is only a very special case of a much larger problem, and even solving that larger problem would not be the final goal. I had wanted to say a little more about each question, but I learnt that the OpenAI problem dump is coming today, so let’s write down some problems in advance, in no particular order.
Your exam begins now. To get full marks you must answer the full question, but partial credit is available. Outside materials including the total sum of human mathematics may be used. Each question is worth 10 points. Hints are given to some of the problems. Good Luck!
-
Lehmer’s Problem I: Does there exist a prime \(p\) such that \(\tau(p)=0\), where
\[q \prod_{n=1}^{\infty} (1-q^n)^{24} = \sum_{n=1}^{\infty} \tau(n) q^n?\]
Although this is a classical problem, it’s hard to understand how interesting it is. Famously, we know that \(\tau(p)/p^{11/2} \in [-2,2]\) by Deligne, and we now know that these values are distributed according to the Sato–Tate measure. On probabilistic grounds, \(\tau(p)=0\) should occur only finitely often; indeed, even \(\tau(p) \equiv 0 \bmod p^2\) should occur only finitely often. So after checking the first billion primes, the answer is presumably no. But the point is not so much the answer itself as what the theory of modular forms can tell us about it. We don’t even know what type of question this is! One can relate it to a question about how many ways \(p\) can be written as a sum of \(24\) squares, but that hardly seems promising. Much more promising, to me, is the fact that the associated Galois representation
\[\rho: G_{\mathbf{Q}} \rightarrow \mathrm{GL}_2(\mathbf{Q}_p)\]
has large global image, while \(\tau(p)=0\) would force it to be locally induced at \(p\). Now the question starts to connect to other conjectures, such as Greenberg’s conjecture, which says that, in this setting, if the Galois representation were locally split, then \(\rho\) would have to be CM (still open!). But this relationship is undercut by the fact that there are examples of modular forms \(f\) with \(a_p=0\) that are not CM, most notably when \(f=f_E\) has weight \(2\) and \(E\) is non-CM but supersingular at \(p \ge 5\). Now weight \(k=2\) is different on probabilistic grounds, but there are also examples with \(a_p=0\) in weights \(k=4\) and \(k=6\).
One can also ask about the analogue for higher-weight forms of level \(N=1\). Now the question starts to relate to Maeda’s conjecture, which predicts that, for a normalized cuspidal eigenform of weight \(k\), the field \(K\) generated by its coefficients has degree equal to the full dimension \(d\) of the space of cusp forms. Stronger versions predict not only that the Galois group of the Galois closure is \(S_d\), but also that \(K\) is generated by any single coefficient \(a_p\) of \(f\). The latter implies that \(a_p \ne 0\) as soon as the dimension of the space is greater than one. Again, it’s totally unclear whether the conjecture is false, true for probabilistic reasons, or true for some good reason.
-
Lehmer’s Problem II: Is the Mahler measure of a nonzero algebraic integer that is not a root of unity bounded away from \(1\)? Progress here would be really interesting. One aspect of this problem is that I don’t think we have a convincing heuristic either way. There has been some nice progress in this general area recently (by Dimitrov, Smith, Orloski–Sardari…), although these methods do not seem to have much to say about Lehmer’s problem. A proof would probably be more interesting than a family of counterexamples, but who knows.
-
Artin’s conjecture for \(2\)-dimensional Galois representations. Naturally, one can restrict to the case of even representations with projective image \(A_5\). If one can answer this, then one can also ask about the full Artin conjecture for any nontrivial irreducible Artin representation of \(G_{\mathbf{Q}}\).
-
The modularity of elliptic curves over general number fields. The case of imaginary quadratic fields seems very much like something that could be done, although I don’t know how to do it. But general number fields are more interesting. Here one can also ask about—and might have to solve—the problem of constructing Galois representations for automorphic forms of algebraic type, as explained in Buzzard–Gee. One can further ask for the modularity of all compatible systems of Galois representations, which encompasses both this question and the previous one on Artin’s conjecture.
-
Leopoldt’s conjecture. It used to be said that you were not a real algebraic number theorist until you had proved Leopoldt’s conjecture (usually incorrectly). More generally, for geometric Galois representations \(V\), there are conjectures about \(H^1(\mathbf{Q},V)\) and the dimension of its image in \(H^1(\mathbf{Q}_p,V)\), including Janssen’s conjecture.
-
The algebraic independence of \(\zeta(2),\zeta(3),\zeta(5),\zeta(7),\ldots\). More generally, the Grothendieck period conjecture (even just for mixed Tate motives!).
-
The existence of infinitely many ordinary primes for any cuspidal modular eigenform. More generally, a Newton = Hodge theorem for a density-one set of primes for any motive \(M\) over any number field \(F\) whose Sato–Tate group is connected.
-
The BSD conjecture in its full form, in arbitrary rank. The BSD conjecture elliptic curves over any number field. Then Beilinson’s conjecture…
-
A disproof of Vandiver’s conjecture. It’s surely false (?!), so this is more a question of finding some really clever way to compute \(K_{12}(\mathbf{Z}),K_{16}(\mathbf{Z}),K_{20}(\mathbf{Z}),\ldots\) until the order of one of these groups has an interesting prime factor. More generally, prove that there are infinitely many primes \(p\) for which it fails. At the same time, prove there are infinitely many regular primes.
-
Last but not least, RH, or GRH, or GRH for \(L(s,M)\) attached to any irreducible motive \(M\).
One could go on and on, for extra credit: Schanuel’s conjecture. Given an algebraic curve \(X/\mathbf{Q}\), is there an algorithm to determine \(X(\mathbf{Q})\)? Are there infinitely many Mersenne primes and finitely many Fermat primes? What comes after the Langlands conjectures? …
I don’t know how well ChatGPT will do, but 10/100 would be extremely impressive to say the least. I would certainly guess that Vandivier’s conjecture is the easiest to disprove since it is “only” a computation. But it might be quite a difficult computation, I’m not sure. And even knowing ALL of these wouldn’t answer many questions I am still interested in.
Update: The results are in here. As BLH says, modularity of elliptic curves over imaginary quadratic fields is in, for 2 points. The quasi-Riemann hypotheses (!?) for 3 points (which proves Artin’s conjecture, but not the Artin’s conjecture I gave above!). So that’s on the one hand 5/100 but on the other hand still OMFG Jaw-droppingly amazing (and of course who knows how well it would have done had it been given the list of questions…)
Why do you think Vandivir’s conjecture is surely false? The best evidence that I know of against its truth are:
1. Washington’s heuristic argument in his book that counterexamples should be extremely extremely sparse, so our computational evidence doesn’t mean much (though wikipedia tells me that refined versions of this argument tend to suggest it is true anyway)
2. I hear people keep trying and failing to generalize it to function fields
But since 1 is not so convincing, we’re really only left with 2. But what makes Vandivir’s conjecture over number fields special, in my opinion, has always been that its a consequence of Greenberg’s conjecture (or its generalization), which to me seem like they are on much sturdier ground. And I don’t think Greenberg’s conjecture has a natural generalization to function fields, so I don’t see item 2 as evidence either.
I’m not sure why #1 is not convincing. Also, true conjectures often have natural generalizations (with extra ramification, for example), and this is some very specific conjecture for \(\mathbf{Q}(\zeta_p)\) that fails for \(F(\zeta_p)\) and general number fields \(F\) in general. I don’t know what conjecture of Greenberg you are referring to, but statements like pseudo-nullity are consequences of Vandiver’s conjecture and are strictly weaker.
For some reason I had picked up the false belief that the direction of implication for pseudo-nullity and vandivir’s conjecture went the other way. (I think I was thinking that \varprojlim Cl(K_n)_p had projective dimension 1 over Lambda). So I think I agree with you now
Modularity of elliptic curves over imaginary quadratic field was indeed in the dump. In fact (according to the machine overlord) the method also shows modularity over totally real and (seemingly) CM fields too. No general number fields yet though!
Is this in the dump (I couldn’t find it), or are you combining the dump with things in your brain?
https://github.com/openai/math/tree/main/preprints/Modularity-of-elliptic-curves-over-imaginary-quadratic-fields-October-4-2026
Yes I read that. Where does it say anything about modularity over totally real fields? Sorry for being stupid.
I think BLH is maybe claiming that the same argument should just work
@DH: It’s not literally in the dump, but looking at the new chaining argument it clearly seemed powerful enough to imply the general totally real case. SO I tried a quick button push, and it generates some plausible slop confirming this (and also the CM case not containing \zeta_5). I am still reflecting on what it’s doing though!
It seems like both the BSD and Hodge for CM as well as other important conjectures are not formalized. Did they forget to upload them. While LLMs are clearly far more reliable they still make quite a few mistakes. It seems strange.
By far the craziest thing for me (but probably for many people) is quasi-RH and no Landau-Siegel zeros. The argument looks completely demented – it seems to fundamentally use cubic theta functions?!
It’s also amusing to think about which problems they surely attempted but did not succeed with (e.g. I would bet money they tried for infinitude of real quadratic fields with class number one).
I hope you give a longer autopsy of the problem dump. Your recent posts have been very enjoyable (and an important corrective to the overall gloomy tone of many public discussions of this stuff).
As an graduate student i am stuck between using AI in my research and not being left behind, and the fact that I will risk never publishing if I use it. And this technology is new so that I feel no one ( including PhD supervisors) has a near-future outlook.
Could you specify which Artin conjecture follows from the quasi-RH?
Actually the primitive root conjecture should follow from quasi-GRH (a gut feeling from how that argument goes) but that would need RH for some more general L-functions that they don’t prove, so the proof of the primitive root conjecture is only in a weak form not in the correct Hooley asymptotic version. One way to try to prove the primitive root conjecture (say for \(a=2\)) is to construct primes \(p = 8n+3\) where \(p-1 = 2q\) for a prime \(q\). Finding such primes is close to a twin-prime sort of problem, of course. But instead you could find *lots* of primes \(p=8n+3\) where \(p-1=2qa\) and where the prime factors of \(a\) are controlled enough (say of size \(p^\eta\) for \(\eta \in [a,b]\) and a,b strictly between 0 and 1 where *lots* means something like \(x/(\log x)^2\), and then the other parts of Hooley should be good enough (not sure how precise one has to choose \(a\) or \(b\)). This is still hard, but this is I think maybe much less hard once one has quasi-RH for Dirichlet L-functions.
Pingback: Updates: Sharing AI progress on mathematics (amazing!); and my lecture plans | Combinatorics and more
A thought experiment I don’t have an answer for:
What would happen if a state of the art AI model was trained only on mathematics up through the 19th century, and you asked it to prove Fermat’s Last Theorem? I think it’s plausible that the existing models could do it, but the proof would be an impenetrable 10,000-page wall of equations. I’m not entirely convinced it would reinvent 20th century number theory.
On the other hand, I’m also not convinced a different set of *humans* would reinvent 20th century number theory.