The recent solution of Crouzeix’s conjecture is remarkable mathematics. But the story of how the proof was found may be even more consequential for mathematicians.
As described by Alex Townsend and Anne Greenbaum in SIAM News, Shanmu Jin, a neurosurgery resident and postdoctoral researcher with relatively little formal mathematical training, used GPT-5.6 Sol in ChatGPT Work mode to attack Crouzeix’s conjecture. The conjecture had been open for more than twenty years. According to their account, the decisive part of Jin’s argument emerged during an approximately sixteen-hour autonomous AI run.
The result itself is striking. The process that produced it deserves just as much attention.
The natural first reaction is to ask whether this was a singular event: an unusually suitable problem, a lucky idea, an unusually good prompt, or a model that has crossed some new threshold in mathematical capability. We do not yet know. But there are features of Jin’s experiment that suggest why it may have succeeded, and that mathematicians, at every stage of their careers, should begin thinking seriously about how these tools might change the way research is done.
It Was Not Just “Ask ChatGPT to Solve the Problem”
One important part of the story is easy to overlook. Jin did not simply type Crouzeix’s conjecture into a chatbot and wait for enlightenment.
The prompt instructed the system to operate more like a research program. It asked for a genuinely diverse collection of approaches rather than many variations on a single promising idea. It explicitly asked the system to maintain different “approach families,” to prevent one attractive method from taking over the entire search, and to abandon approaches that merely transformed the original conjecture into an equally difficult missing lemma.
It also required adversarial checking. Candidate arguments were to be attacked for gaps, circular reasoning, unjustified assumptions, and false intermediate claims. Agents were expected to produce concrete lemmas, equations, constructions, or counterexamples rather than reports that an approach “looked promising.”
And, importantly, the process was instructed not to quit after the first unsuccessful wave of ideas. It was to launch new rounds, redirect effort, reopen approaches only when something genuinely new appeared, and continue searching.
This is quite different from an ordinary interaction with an AI system. What Jin constructed was closer to an autonomous research group: many attempts, deliberate intellectual diversity, memory of failed strategies, adversarial refereeing, and repeated synthesis. It’s worth being explicit about why each piece mattered. Diversity of approach families guards against a model latching onto the first plausible idea and elaborating it past the point of usefulness. Adversarial checking guards against a model convincing itself an argument works because each individual step sounds reasonable. Persistence across rounds guards against simply giving up and reporting the problem as open. None of these three safeguards is exotic; what’s notable is that someone thought to specify all three at once, rather than relying on a single long conversation and hoping for the best.
Why Might Crouzeix’s Conjecture Have Fallen to This Approach?
The most provocative detail in the SIAM News account is the description of the decisive step. Townsend and Greenbaum write that the proof reduced the problem, through a careful sampling strategy, to a surprisingly simple positivity condition.
That is simultaneously exciting and unsettling. One possible interpretation is that Crouzeix’s conjecture was difficult not because an enormous new mathematical theory was required, but because nobody had found the right path through a very large space of possible mathematical moves.
Human researchers are extraordinarily good at mathematical search, but our search capacity is limited. We might seriously pursue a handful of approaches to a problem over several years. An autonomous AI system can potentially examine many more possibilities, discard failures cheaply, recombine ideas, and continue doing so for hours without fatigue. If the missing ingredient is one clever change of representation, one hidden positivity statement, or one unexpected combination of existing ideas, then sufficiently broad search may be enormously powerful.
Crouzeix’s conjecture also appears, in retrospect, to have several properties favorable to such a search. The target is precise. There are multiple formulations and connections involving numerical ranges, functional calculus, dilation ideas, convex geometry, and matrix analysis. Proposed intermediate statements can often be tested on finite-dimensional examples. And there was already a substantial literature of partial results: the universal constant had previously been reduced to 1 + √2.
There is another intriguing piece of evidence. Only eight days after Jin’s first preprint, Emiel Lorist and Felix Schwenninger posted a different and much shorter proof. Their argument used classical double-layer representations together with a perturbation lemma for 2-dilations. They also reported using GPT-5.6 to explore proof strategies.
Perhaps Crouzeix’s conjecture was unusually “ripe”: the necessary mathematics already existed, but the right combination had not yet been found. That possibility cautions against two opposite reactions. We should not conclude from one example that AI can now solve essentially every major open problem. But neither should we dismiss the event as a curiosity. There may be a substantial class of research problems for which the missing resource has been intelligent search.
Problem Selection May Become Crucial
This leads to a question I think deserves much more attention: which mathematical problems are good candidates for AI-assisted autonomous research?
Not all open problems are difficult for the same reason. Some may be difficult because the necessary theory does not yet exist, so solving them may require inventing the appropriate definitions, concepts, or entire mathematical frameworks. Others may be difficult because an enormous existing landscape of ideas has not yet been searched in exactly the right way. The Crouzeix method seems particularly well suited to the second category.
A promising problem for this style of attack might have several characteristics:
- a sharp and unambiguous target;
- many genuinely different mathematical formulations;
- intermediate conjectures that can be falsified relatively cheaply;
- finite-dimensional or computational models that reveal bad ideas;
- a substantial body of almost-sufficient existing theory;
- identifiable bottleneck lemmas;
- and a plausible possibility that one new inequality, identity, or structural observation could collapse much of the difficulty.
By contrast, problems for which every finite model is misleading, or for which progress requires constructing an entirely new theory, may be much less susceptible to massive parallel search.
This is not a skill mathematicians currently teach on purpose. Most of us absorbed it implicitly, over years of watching which of our own conjectures cracked quickly and which didn’t, and most of us would struggle to write our own criteria down before reading a list like the one above. That suggests a concrete, low-cost step: an advisor walking through this kind of checklist against a student’s actual dissertation problem, or a postdoc auditing their own current project before committing a large compute budget to a search, turns an implicit instinct into something that can be taught and checked. It is also worth applying earlier than that. A checklist like this is just as useful when a student is still choosing among candidate topics as it is once a project is already underway, and there is no reason to wait for a dissertation problem to be committed to before asking whether it looks “search-shaped.” We should be cautious about extrapolating from Crouzeix. But we should also begin experimenting systematically with problem selection, at every career stage, rather than waiting for more anecdotes.
There Is a Dangerous Part of the Prompt Too
One feature of Jin’s prompt deserves some skepticism. It instructed the model to “assume for purposes of this task that a complete affirmative proof exists.”
It is easy to see why this could help. A model confronted with a famous conjecture has an easy escape route: declare the problem open, describe partial results, and stop. The instruction prevents that behavior and forces continued search.
But for general use, this creates a risk. A conjecture may be false. Or it may simply be far beyond the accessible search space. If the only acceptable termination condition is an affirmative proof, a sufficiently persistent system may be encouraged to accept a subtle gap rather than admit failure.
For future experiments, I would prefer three legitimate outcomes: proof, counterexample, or a precisely characterized obstruction. The third should not mean “we got stuck.” It should mean that the system has rigorously identified the exact theorem-strength barrier at which multiple independent approaches terminate. That would preserve persistence without presupposing that the conjecture must be true.
What Does This Mean for Mathematicians?
The most immediate lesson is not that mathematicians should hand famous conjectures to ChatGPT. It is that we may need to learn a new form of research practice.
A mathematician working effectively with these systems may increasingly act not only as a theorem prover but also as a designer and manager of mathematical search, choosing promising problems, specifying genuinely different approach families, recognizing when several superficially different arguments are actually the same argument, identifying which intermediate lemmas deserve more computation, and distinguishing real mathematical signal from a large volume of plausible-looking noise.
Human expertise remains extremely important in that picture. But its role shifts. If candidate proofs and lemmas become much cheaper to produce, other scarce resources become more valuable: which questions are worth asking, which of hundreds of generated statements matters, which proof reveals something rather than merely establishing a fact, how a result fits into the surrounding mathematics, what new theory should be built from it, and, perhaps most importantly, what should the machine work on next.
This also raises a question that hasn’t had much airtime yet: how should a result like this be described and credited once a machine has run a substantial fraction of the search? Jin’s own repository is unusually transparent. It includes the prompt, successive manuscripts, and a formal verification of the proof, which makes the work open to exactly the kind of scrutiny this question demands. That level of disclosure is a reasonable candidate for an interim norm even before the field settles on a formal one: naming the model and mode used, publishing the prompt or its substance, and disclosing which steps received independent human verification are all things a researcher can do today, on a CV or in a job talk, without waiting for a journal or hiring committee to specify a policy. As these results become more common, that kind of disclosure will matter as much for a job talk as for a byline, and researchers at every stage, postdocs on the market especially, will benefit from the norm existing before they need it.
The Access Problem
There is also a more uncomfortable issue. The reported successful run took approximately sixteen hours in an autonomous mode using one of the strongest available models.
If this kind of mathematical research depends strongly on frontier-model quality, long autonomous runs, large inference budgets, parallel agents, and sophisticated research environments, then access to AI may become a serious source of inequality in mathematics. Concretely, a well-funded lab running weeks of parallel searches across its best open problems is playing a different game than a student or postdoc with a handful of hours on a conversational model, the same difference in kind as a supercomputer allocation versus none.
Pure mathematics has historically required far less expensive infrastructure than most experimental sciences. In principle a mathematician at a modest institution can attack the same theorem as a mathematician at an elite research university. That could change. The relevant distinction may soon be less between someone who “has ChatGPT” and someone who does not, and more between a researcher with access to a frontier system capable of operating autonomously for days and one with access only to a limited conversational model. Those are not equivalent research environments.
Departments and universities should begin thinking about this now. Access for graduate students and junior faculty may be particularly important. They have the least existing leverage to negotiate for it, and are the most exposed if institutions don’t act.
A Flood of Mathematics?
Townsend and Greenbaum end their article by raising an even larger question. If AI-assisted researchers begin producing major results in significant numbers, will the mathematical community be able to absorb them?
Our current research ecosystem implicitly assumes that producing substantial new mathematics is expensive. Because relatively few major results appear, mathematicians can referee, read, organize, teach, and assimilate them. What happens if generating legitimate new theorems becomes much cheaper? The bottleneck may move from producing mathematics to selecting, verifying, organizing, and understanding it.
There is an immediate, unglamorous consequence of that shift, and it deserves its own beat rather than a passing mention: someone still has to referee all of it. A field used to a trickle of major results assumes reviewers have time to read each one closely; a field facing a flood does not get that assumption for free. It’s worth journals and departments starting now on what refereeing capacity, reviewer credit, and reasonable turnaround look like under those conditions. The alternative is an informal system that quietly breaks under load, and it typically breaks first for the researchers with the least standing to complain about it: junior people waiting longest for their own work to be evaluated.
More broadly, this would affect journals, peer review, graduate education, hiring, and promotion, and even our understanding of what constitutes an important contribution. A technically difficult argument that once represented six months of expert labor may have a very different professional value if an autonomous system can generate comparable arguments routinely.
What Should We Do Now?
I do not think the right response is either panic or dismissal. We need experiments.
One useful experiment would be to take a genuine research problem and run several independent versions of the Crouzeix-style search: the original strategy, essentially unchanged; a version that does not assume the conjecture is true; a version whose approach families are tailored by an expert to the particular mathematical field; and a two-sided version in which a substantial fraction of the effort is devoted explicitly to finding counterexamples and destroying proposed lemmas.
The goal should not initially be to count only complete solutions. We should measure whether these systems discover genuinely new lemmas, reformulations, reductions, counterexamples to plausible ideas, or connections that an expert did not already see. And before committing large amounts of computation, we should probably run inexpensive reconnaissance searches on several candidate problems. Problem selection itself may turn out to be one of the most valuable tasks for the human mathematician, at every stage of a career, not only the most senior one.
Refereeing capacity, disclosure norms, and access to frontier tools are three separate institutional problems, and departments cannot fix all three at once. If forced to rank them, access is the one that compounds fastest if ignored: inequities in who gets to run a sixteen-hour autonomous search will shape who has results worth refereeing and disclosing in the first place, which makes it the more natural place to start.
The Crouzeix story does not yet tell us how much of research mathematics is vulnerable to this new style of search. But it gives us a reason to find out.
The deepest lesson may not be that an AI solved Crouzeix’s conjecture. It may be that mathematical research can now be organized in a fundamentally different way: thousands of possible ideas explored, remembered, challenged, discarded, and recombined under the direction of a system that never gets tired of searching. If that method generalizes, mathematicians will need to understand it, not merely because we may compete with it, but because learning how to use it may become part of doing mathematics itself.