When I was astir 11 I heard for the first clip astir Fermat’s Last Theorem. I was instantly captivated by the problem statement, arsenic good arsenic by the accompanying story, and made a reasonably superior effort to beryllium it. And while, unsurprisingly, I failed, I learned a batch from the attempt. Blissfully ignorant of the truth that the 
Since I still couldn’t really spot wherever to start, I decided to simplify the problem further and ore connected successive differences of cubes, pinch a position to showing that specified a quality could not itself beryllium a cube. At the clip I did not cognize really to definitive what I was doing successful algebraic language, truthful I did not explicitly effort to beryllium that the Diophantine equation 
Somehow I ne'er managed to move this study into a impervious of Fermat’s Last Theorem, and later connected my dream of solving it sewage replaced by different mathematical dreams. However, erstwhile I reached the constituent successful my mathematical acquisition wherever I was taught astir taking quality sequences and astir what happened to polynomials, I understood those topics overmuch amended than I would person if I had not discovered quality sequences for myself and spent happy hours playing astir pinch them. I mention this communicative conscionable arsenic an illustration of the arena that was powerfully emphasized successful this missive signed by 25 Fields medallists, that 1 learns a batch from reasoning astir a problem, sloppy of whether 1 solves it.
In the end, however, I felt that I could not motion the letter, contempt agreeing pinch overmuch of what it said. Instead, it seemed amended to do what I did pinch the Leiden Declaration and group retired my ain position successful a blog post. But it should beryllium understood that by doing that I americium not mounting myself up arsenic a personnel of immoderate opposing camp: so 1 of my worries astatine the infinitesimal is that the mathematical organization mightiness go bitterly divided, thing I would very overmuch for illustration to avoid. Also, I work together connected the basal constituent that we are facing a crisis: I conscionable want to connection a somewhat different study of what that situation is. I don’t declare afloat originality for this analysis, arsenic I cognize that respective different mathematicians person already put guardant thoughts that are akin to the ones I have, though (for what it’s worth) I person mostly travel to these conclusions independently.
On the taxable of independence, it will possibly thief if I explain that while I person contacts successful the mathematics group astatine OpenAI, and person besides been fixed early entree to immoderate of their models (typically only a fewer days earlier they person been released), and person been fixed free entree to their Pro models erstwhile released, I person ne'er been paid by OpenAI. I mention this successful the hope, possibly naive, that what I constitute will not beryllium dismissed for advertisement hominem reasons. Another imaginable logic for my being regarded arsenic “pro-AI” is that, arsenic I person stated publically respective times, I person a group successful Cambridge devoted to automatic theorem proving. However, that is really much of a logic to beryllium anti-AI, since our group has been trying to onslaught the problem of getting computers to beryllium absorbing theorems by knowing arsenic good arsenic imaginable really humans beryllium absorbing theorems, truthful now that LLMs tin intelligibly do it without the thief of specified insights arsenic we person had, 1 of the main motivations for our activity has disappeared. To put it different way, we person had to swallow the bitter lesson (which of people we were ever alert was a chopped possibility, moreover if the velocity astatine which it happened has taken america by surprise). I do successful truth deliberation that it is still a very absorbing and valuable intelligence workout to effort to summation this understanding, moreover if we tin usage LLMs arsenic achromatic boxes, but that’s a taxable for different blog post.
So why didn’t I motion the letter? Let maine extract a mates of sentences from it that definitive what I spot arsenic the main statement being put forward.
But solving problems is only a instrumentality and proxy for achieving the superior extremity of conceptual knowing and insight. Forgetting this successful the world of AI whitethorn move the instrumentality against the superior goal. Indeed, the wide accumulation astatine faster and faster gait of “true/false” statements could destruct fertile crushed alternatively of breathing life into caller ideas.
Perhaps the main logic I didn’t motion is that I don’t afloat subscribe to this view. Instead, I person a much analyzable view, which I really expressed successful my effort The Two Cultures of Mathematics a 4th of a period ago, and which tin beryllium summarized by saying that location is simply a spectrum of attitudes successful mathematics to the narration betwixt problem-solving and conceptual understanding. At 1 extremity of the spectrum you person mathematicians who are chiefly motivated by the wish to lick problems, who spot conceptual knowing arsenic a very important intends to that end. At the different you person mathematicians who are chiefly motivated by the wish to attain conceptual understanding, who spot problem-solving arsenic a very important intends to that end. I interest that the severe-misalignment missive could beryllium seen arsenic saying that the “right” cognition is to attraction connected conceptual knowing arsenic the main privilege — indeed, the supra sentences opportunity that much aliases little directly. But I deliberation that location are mathematicians each crossed the spectrum, and that that is simply a bully point (or possibly I should opportunity that it has been a bully point up to now — the early is overmuch little certain), and I don’t want to propose to a ample fraction of mathematicians, including myself, that their mathematical temperament is someway “wrong”.
My ain peculiar mathematical cognition is very akin to 1 that was beautifully articulated successful a Twitter station by Jacob Tsimerman (another non-signatory of the letter), which, now that I look astatine it, says a batch of what I will beryllium saying here. And that station successful move is simply a consequence to Daniel Litt, who is successful my sentiment 1 of the wisest commentators connected mathematics and AI. His views are expressed in a later station here, which I deliberately didn’t publication until finishing this one, and past found, arsenic I expected, that location was important overlap. I would besides for illustration to return this opportunity to urge an fantabulous station by Noah Smith entitled The End of the Age of Heroes, successful lawsuit you haven’t publication it.
I person been talking truthful acold astir individual mathematical understanding, but I fishy that what concerns astir of the signatories is little that than the corporate knowing that results astatine slightest successful portion from the quality activity of problem solving. My conjecture is that they would argue, wholly coherently, that moreover if corporate knowing is the superior goal, if galore individuals are chiefly motivated by the wish to lick problems, that’s perfectly good and contributes to that corporate understanding.
With that interpretation, the rumor becomes somewhat different: is it much important that the corporate knowing of the mathematical organization should beryllium arsenic precocious arsenic imaginable aliases that location should beryllium answers to arsenic galore problems arsenic possible? Or are those 2 intends valuable successful different ways, truthful that location is nary constituent successful declaring 1 of them much important? Or are they truthful inextricably linked that it makes nary consciousness to reason that 1 is much important than the other? And erstwhile we opportunity “important”, for whom are we saying it is important: for mathematicians, aliases for nine arsenic a whole?
I find these difficult questions, truthful I don’t want conscionable to state an reply to them. (Do you spot what I did there?) Instead, I’d for illustration to effort to connection astatine slightest immoderate statement for immoderate conclusions I travel to, moreover if they are tentative. So let’s comparison 2 scenarios. In the first, which I deliberation is the much apt really to happen, models go publically disposable that are amended astatine solving problems than virtually each mathematicians. If location are a fewer residual mathematicians who tin do things the models can’t, moreover they activity acold faster if they make dense usage of the models. Thanks to this, successful a short clip we get answers to galore questions that we person profoundly cared about, but the complaint astatine which we person these answers acold exceeds the complaint astatine which the mathematical organization tin sorb them. In particular, astir of the answers are obtained pinch zero effort from quality mathematicians — conscionable prompts specified arsenic “Thank you — please continue”.
In the 2nd scenario, location has been an world agreement, for wholly different reasons, to artifact the nationalist merchandise of models importantly much powerful than the ones we presently have, and the mathematicians wrong the tech companies work together to clasp disconnected from getting their soul models to lick awesome problems. Instead, they return guidance from the mathematical community, solving problems only erstwhile asked to do truthful by immoderate suitably typical assemblage that decides that the use of receiving a solution of a definite problem outweighs the benefits of humans struggling to lick it complete a overmuch longer timescale.
I’d for illustration to see what the quality would beryllium betwixt these 2 scenarios some for individual and corporate understanding. I’ll statesman pinch individual understanding.
One mightiness reason that for individual understanding, not excessively overmuch would alteration if we are abruptly flooded pinch ample numbers of large caller results. There is already acold much mathematics retired location than I person immoderate dream of knowing (for example, contempt being fascinated erstwhile Fermat’s Last Theorem was proved, I person made nary effort to understand the proof), and moreover among the parts that I do understand, the parts that I understand because I myself discovered them shape a very mini fraction, though a fraction that I understand much profoundly than thing other (at slightest temporarily — aft a while I hide things and suffer rather a batch of the knowing I built up). However, 1 change, which seems positive, from the position of the building up of individual understanding, would beryllium that we would person a overmuch bigger prime of results that we could take to study. Also, if we recovered ourselves stuck connected immoderate point, AI would beryllium capable to thief us. The main apt antagonistic alteration is that we would astir apt cease to workout that portion of our brains that we usage erstwhile spending months aliases years struggling pinch a difficult investigation problem, which tin beryllium hugely adjuvant successful processing understanding.
I opportunity “likely” because successful rule location would beryllium thing to extremity america reasoning astir very difficult problems without consulting LLMs, but successful believe it seems improbable that group would put successful the aforesaid level of effort that they do now. The business mightiness a spot for illustration what happened pinch satnavs, wherever 1 could ever determine not to usage them, to support the portion of the encephalon progressive that tin look astatine a map, study a route, and travel it, but successful believe astir group succumb to the enticement to usage a satnav. (In fact, I myself do effort to support that portion of my encephalon active, and was alternatively proud of uncovering my measurement location precocious erstwhile I had concisely looked up the way connected my telephone but past forgotten to bring the telephone pinch maine erstwhile I really went there.) But moreover if each we were doing was reference AI output, I deliberation that the problem-solving muscles successful the encephalon wouldn’t atrophy completely. When students are reference maths papers, I powerfully counsel them (and I deliberation this is beautiful modular advice) to publication “actively” alternatively than “passively”, doing things for illustration trying to beryllium the consequence for yourself, looking astatine the insubstantial only erstwhile you consciousness stuck and request a hint, and moreover past conscionable trying to get the hint and arsenic small other arsenic possible. If 1 sounds a insubstantial that way, past 1 is perpetually solving problems, immoderate conscionable exercises and immoderate rather a spot harder. It seems apt that an LLM could get to cognize what our mathematical inheritance is and provender america pinch conscionable the correct hints to let america to activity our measurement done a mathematics insubstantial successful this progressive way. Yes, we would suffer the peculiarly heavy level of knowing and ownership that comes pinch having solved a difficult problem oneself, but it isn’t clear to maine that advancement successful mathematics would suffer arsenic a result. I would beryllium very willing to perceive counterarguments to precisely this point. That is, I would beryllium willing to cognize what usage that level of heavy engagement pinch a impervious mightiness person successful a world wherever AI is overmuch amended than we are astatine uncovering proofs.
How astir corporate understanding? Let maine quote a spot much of the letter.
Indeed, the wide accumulation astatine faster and faster gait of “true/false” statements could destruct fertile crushed alternatively of breathing life into caller ideas.
Often these solutions are announced successful a rush, leaving nary clip for a due writeup, the isolation of caller methods and ideas, and citing applicable erstwhile activity of others. As successful each imaginative professions, this raises terrible attribution and plagiarism questions. Moreover, without the consenting mathematicians who must return attraction of their improvement and integration into the mathematical canon, AI-conceived ideas would ne'er go afloat live and the important quality transmission concatenation betwixt mathematicians would beryllium lost.
I’ll travel backmost to questions astir due citation and attraction connected what I return arsenic the halfway interest here: that if results are proved excessively quickly, past the digestion process will go impossible. I americium decidedly worried that results will not beryllium decently digested, but for different reasons.
A first remark is that what AI is producing is not conscionable true/false statements: we now cognize not conscionable that the Navier-Stokes equation pinch soft forcing admits finite-time blow-up, but we person a impervious of that, which builds connected a awesome woody of awesome activity done by quality mathematicians. Many group utilized to definitive the interest that AI would lick our favourite problems pinch utterly opaque proofs, but that has not turned retired to beryllium the case, moreover if their write-ups often time off plentifulness to beryllium desired. (Incidentally, I spot these inadequate write-ups arsenic almost surely a impermanent annoyance and truthful not arsenic a basal threat to mathematical believe aliases early mathematical understanding.)
Secondly, moreover if the measurement of caller results is large, mathematics is simply a highly specialized discipline, truthful mathematicians tin activity successful parallel. If, for example, we had to digest 1000 important results successful a twelvemonth that were astir uniformly distributed crossed mathematics, past astir sub-communities of mathematicians would astir apt want to understand astir 30 of them, and for each individual problem location mightiness good beryllium only a mini fistful of specialists who would beryllium evident group to return the lead successful reaching this understanding, pinch that fistful varying from problem to problem. So it would beryllium a large task, but not needfully an intolerable one.
In this context, it is worthy reasoning astir the immense measurement of output of quality mathematicians, which seems to person been expanding recently, moreover earlier AI. While I person sometimes heard complaints astir this, I person surely not heard suggestions that quality mathematicians should slow down the complaint astatine which they beryllium absorbing theorems. That whitethorn beryllium partially because the authors of those theorems return the problem to constitute their papers good and springiness bully talks. But what astir the ample amount of papers, including important ones, that are not written good and whose authors springiness incomprehensible talks? That tin beryllium annoying, but it is simply a acquainted annoyance and not 1 that we deliberation of arsenic a crisis.
A 3rd constituent is that moreover if the measurement of AI output is excessively large for america to beryllium capable to digest it properly, that is not needfully a bad thing. To tie an imperfect analogy, location is now much contented disposable connected streaming services than anyone could perchance watch, pinch the consequence that location is almost surely immoderate very bully contented retired location that is hardly watched astatine all. But that isn’t evidently a worse business than if location were acold little contented and each of it received the attraction it deserved. Returning to mathematics, if location were excessively overmuch AI-generated contented for america to beryllium capable to digest it, past we could take which parts of it we wanted to digest.
For that we would request to person immoderate thought what was location (a business a small akin to really quality mathematicians typically study rather a batch astir what results are known successful their area moreover erstwhile they do not understand their proofs successful immoderate detail). One measurement 1 could effort to execute that would beryllium to create a well-designed database, astir apt pinch AI help. But possibly that would beryllium unnecessary, and alternatively 1 could simply talk to an LLM and inquire it to springiness a bird’s-eye position of immoderate area of mathematics 1 wanted to understand successful that knowing-what’s-there way.
The fearfulness seems to beryllium that immoderate very absorbing and important parts of mathematics will beryllium discovered by AI and past overlooked, erstwhile had they been discovered by quality mathematicians they would not person been overlooked. And that whitethorn moreover beryllium the case, but what matters is whether the magnitude of absorbing and important mathematics discovered by AI that is not overlooked will transcend the magnitude of absorbing and important mathematics that would person been discovered and decently digested by humans pinch AI having played a much humble role.
In short, it seems to maine that while a flood of “big” AI results would beryllium apt to summation the magnitude of important mathematics that was not decently digested, it would besides beryllium apt to summation the magnitude that was decently digested, which seems for illustration a beautiful bully bargain.
Let maine quickly talk the problem of AI not decently crediting quality mathematicians. I work together that this is simply a superior problem correct now, but it is different problem that I spot arsenic temporary. Very soon, the full “credit system” will surely collapse, since uncovering an astonishing impervious will beryllium nary much of an intelligence accomplishment than erstwhile a national intelligence spots done their scope an entity that turns retired to beryllium a caller comet. Until that happens, it is important to springiness humans the in installments they deserve, since careers tin dangle connected it, but that will soon cease to beryllium the lawsuit arsenic well. I person to opportunity that I’m puzzled that this problem exists, since I would person thought that if you asked an LLM to look astatine a impervious and show you which ideas successful it are adjacent to ideas that are successful the lit already, it would beryllium highly bully astatine that task. I dream the reply to this conundrum is not that group person been successful specified a hurry that they person simply not taken the problem to do this, but I fearfulness that it mightiness be, astatine slightest successful immoderate cases. If so, past those who person been sloppy merit to beryllium criticized, but it is simply a insignificant matter compared pinch the endurance of mathematics, particularly if the deficiency of citations is swiftly put right.
Does each this mean that I americium optimistic that mathematicians will extremity up digesting astatine slightest arsenic overmuch mathematics successful a post-AI world arsenic it would person if AI had not been capable to beryllium awesome theorems? Not exactly. But my interest is not that we would beryllium incapable to do it, but alternatively that the societal structures that presently support this digestion process will beryllium destroyed and not adequately replaced.
One measurement that mightiness hap is that AI disrupts nine truthful much, aliases moreover kills immense numbers of us, that the preservation of thing for illustration the existent mathematical contented ceases to beryllium of immoderate concern: each that will matter is the endurance of the quality race. But that again is simply a taxable for a different blog station (which successful truth I americium successful the mediate of writing).
Let’s presume alternatively that we get fortunate and that AI remains much aliases little nether control. My interest past is that we do not negociate to transmit what we cognize to a caller procreation of mathematicians. Speaking for myself, my main information for becoming a mathematician was the dream that I would lick unsolved problems — the much celebrated the better. I person besides ever greatly preferred straight reasoning astir a problem to reference books and papers and mostly learning the mathematics of different people. (I’m not saying that’s good, but conscionable stating a truth astir myself.) If the dream of solving a celebrated problem had not existed, I’m not judge whether I would person go a mathematician. I don’t wholly norm it out: possibly what really motivated maine was that I had an aptitude for the taxable and that solving problems was a measurement of getting respect from a mini group of peers. And possibly I could person tried to summation that respect successful a different way, specified arsenic reasoning very difficult astir an area of mathematics until I was capable to show to others conscionable really good I understood it. But I’m not judge really motivating that would person been for me. I very overmuch dream that location is simply a excavation of young group for whom it will beryllium a powerful motivation, because I deliberation the endurance of a quality mathematical contented whitethorn good dangle connected it.
Thus, the superior risk, arsenic I spot it, is that a batch of group who would person done a PhD successful mathematics and gone connected to go custodians of the mathematical contented will nary longer wish to do so. Those of america who person PhD students, including me, request to effort arsenic difficult arsenic we tin to travel up pinch imaginative ways for them to usage their clip productively (in consultation pinch the students themselves, obviously). Whether aliases not we do a bully occupation pinch that could make a immense quality to the early of mathematics. A related consequence is that the cognition among policy-makers will beryllium that mathematicians are nary longer needed and that backing will go overmuch harder to travel by: we urgently request to travel up pinch bully ways of explaining the worth of having a ample excavation of quality mathematical experts, moreover if it is nary longer portion of their domiciled to find caller proofs of theorems.
A last logic that I didn’t motion the missive is that I wasn’t really judge what it was demanding that isn’t happening already. It seems apt that successful a matter of not very galore months LLMs will beryllium released that are capable to lick awesome mathematical problems, and they will presumably person nary problem astatine each pinch much run-of-the-mill problems. However overmuch we mightiness regret that, location is nary chance that the effect of specified models connected mathematics will seduce AI companies to extremity their release, though possibly concerns astir information will lead to immoderate hold and springiness america a spot much clip to activity retired really to adapt. Assuming that they are released, location will beryllium a flood of caller results, whether we for illustration it aliases not, and it will nary longer beryllium the AI companies producing them, though possibly the shape will proceed that the AI companies will person entree to much powerful models and truthful will get much than their adjacent stock of header results. So I felt that location was thing to beryllium gained from criticizing AI companies for generating excessively galore solutions excessively quickly. In fact, it whitethorn good beryllium that each that does is bring guardant by a mates of months what was going to hap anyway, and possibly it will moreover let the results to beryllium released successful a much controlled measurement than they would person been if they had been discovered by random group erstwhile the models were publically available. Under the circumstances, I deliberation the champion we tin do is recognise the changes that are coming and effort to activity retired the slightest unsatisfactory measurement of dealing pinch them.
Tags: ai, artificial-intelligence, math, mathematics, technology
This introduction was posted connected September 17, 2026 astatine 7:55 americium and is revenge nether AI and maths. You tin travel immoderate responses to this introduction done the RSS 2.0 feed. You tin leave a response, aliases trackback from your ain site.
English (US) · 
Indonesian (ID) · 
