Gresham College Lectures
Gresham College Lectures
The End of Mathematics? The AI Crisis - Alain Goriely
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
This lecture was recorded by Professor Alain Goriely FRS on 10th of November 2026 at Barnard’s Inn Hall, London.
Alain Goriely is a mathematician with broad interests in mathematical methods, mechanics, sciences, and engineering. He is well known for his contributions to dynamical systems, mathematical biology, as well as fundamental and applied mechanics. He is particularly well known for the development of a mathematical theory of biological growth, culminating with his seminal monograph The Mathematics on Mechanics of Biological Growth (2017).
The transcript and downloadable versions of the lecture are available from the Gresham College website: https://www.gresham.ac.uk/watch-now/end-maths
Gresham College has offered free public lectures for over 400 years, thanks to the generosity of our supporters. There are currently over 2,500 lectures free to access. We believe that everyone should have the opportunity to learn from some of the greatest minds. To support Gresham College's mission, please consider making a donation: https://www.gresham.ac.uk/get-involved/support-us/make-donation/donate-today
Website: https://gresham.ac.uk
X: https://x.com/GreshamCollege
Facebook: https://facebook.com/greshamcollege
Instagram: https://instagram.com/greshamcollege
Bluesky: https://bsky.app/profile/greshamcollege.bsky.social
TikTok: https://www.tiktok.com/@greshamcollege
Support Us: https://www.gresham.ac.uk/get-involved/support-us/make-donation/donate-today
So, today is a big day. And I had to interrupt a regular program to broadcast an important message. Mathematics is in crisis. And to understand the depth and consequence of this crisis, I want to tell you a little bit about the history of mathematics, mostly pure mathematics, and its relationship with computers. So I want to start quite a while back, before AI was even a word, way before. And here is my first example, the four-color problem. It goes back to 1852. Francis Guthrie and his brother, the brother was a student of August de Morgan, had the question: how do I colour the map of uh British counties that you see here using only four colours? And he managed to show that it was possible and put the question to Auguste de Morgan can we prove that four colors was always uh sufficient? Like in this transformation. And the conjecture for the longest time was that four colors suffice for any ordinary map. It was a problem in graph theory, and there are algorithms to do that, but there was no proof that it can always be done. And here I show you an example that I just did on uh uh Chat GPT, and to tell you to show you the kind of things you can do now very easily that would take a little bit more work, maybe a day of programming. This is a single prompt with the right information about which algorithm to use, and it produced the result of the four colors for the UK counties. So that question was open until 1976, when Apple and Aken came up with a proof. The only problem with this proof is that it was computer-aided. They managed to reduce the problem to 1834 cases that might be a counterexample, and the computer systematically checked that they were actually not counterexample, and there was no possible counterexample, and the proof was correct. The problem with that is that mathematicians did not did not like that at all. It was very uneasy. Says, what now we're going to do things with computer? Is that actually true? Or can we be sure that the computer didn't make a mistake? Remember, a proof for mathematician is a way to convince an unreasonable person, not a simple argument. And so eventually people accept it. I once went to a library and find the whole printout of the four-color theorem. Eventually, it was proved in 2005 by Werner and Gauthier using what's called a proof assistance. And a proof assistant is actually a computer language, ironically, that's what's called formalize a result. That means you state the result very precisely with all the rules of logic within the computer, within the code. And if the code compiles, if the code turns, it guarantees, it gives you a certificate that the proof is correct. And it's widely considered now as the gold standard. The original one was called COC, as a French program, but the current one, the most used one, is something called lean, that is used to guarantee certified, it's called a lean certificate that a proof is correct. And we'll get back to that uh later. Now, a second interesting example of this type of relationship is the work of uh uh Ed Lawrence. So Ed Lawrence in 1963 came up with a system of three differential equations describing the dynamics. It was reduced from fruit mechanics, but the system was interesting in itself. In an absolutely wonderful paper, explained that the dynamic of this system was very complex, and that might be used to understand complex dynamics that's found in the weather system. And that really gave buzz, that was one of the very first examples, the paradigms for things that would be for chaos. And so if you start at one point, what the equation shows you is that in three dimensions, it converged to this iconic picture of with two butterfly wings and with sensitivity to initial condition and all that, a lot of things that were completely understood by Ed Lawrence and fully understand and develop in the 80s and 90s to the general uh theory of dynamical system and chaos. Beautiful expansion of our horizon, mostly due to people being able to use personal computers to do this type of simulation, discovering a lot of facts, and coming up with general theory for that. Here is also an example. Uh, what probably took Ed Lawrence weeks to do in the 60s, that would have taken maybe days in the 80s and the 90s, maybe an hour, a year ago, was done in three minutes and one prompt with ChatGPT, right? That shows you so mind-blowing the acceleration of this technique, at least when you know how to properly use them. On the pure mathematics side, the question was a little bit different. It says, yes, you show stuff on the computer, but do we really know what is this object? We have a very precise definition of what we call a strange attractor. Can you actually prove that this is a strange attractor? That was a big challenge until 1999, between 1999 and 2000, Warwick Tucker came up with a proof. And the proof used computer in a very interesting and innovative way. What it did is by looking at interval with the dynamic. The problem with a computer simulation is that you have only a finite number of digits and that you keep making mistakes on the last digit. You can't keep the whole information. And what uh Warwick did was to show that you can always bond a solution between two intervals using an interval arithmetic and prove that the solution satisfies the condition of the strange attractor, and came up with a rigorous proof that the famous Lorentz attractor was indeed a strange attractor. But by and large, these were mostly considered exceptions. Yes, you know, you can use this computer, it is a really edge case. And the state of affair that's uh well described in this uh little book, Wonderful Little Books, by Tim Gowers. Talk about Tim Gowers, a fit medallist and great commentator about the state of mathematics. There's a mathematics, a very short edition, published in 2006. And in there, he write that if you stuck and need a bright idea, then a computer will be no help whatsoever. In fact, most mathematicians would say the most important tools, piece of paper and something to write with. And that was essentially what was understood maybe until one or two years ago. So what changed? Well, AI came in the picture. And to tell you the story of AI, you know, it grew over the last few years. I'm going to use it through Shakespeare's All the World's a Stage, in which mathematicians are merely players. They have their exits and their entrances, and one machine in its time plays many parts. The first one, the infant, muling and puking. And I'm bringing you back to something that someday that you probably remember because it made the news. November 2022, four years ago, Chat GPT launches. Everybody talks about it, or you can ask the machine something, it replies to you and all that. It's wonderful. You can it's you can make little poetry and everybody giggles, that was great. And says, but can it do math? Well, let's try. In December 2022, for instance, he says, question, is the product of two prime numbers always prime? And Chat GPT, with its normal confidence, replied, Yes, of course, the product of two prime numbers can be either prime or composite. For example, the product of the prime number three and five is 15, which is itself a prime number. Okay, no need to tell you that mathematician did not feel threatened at that point in time. It was all fine. But it didn't last. The whining schoolboy came. Our little baby is growing and growing fast. Six months later, DeepMind came up with a program, Alpha Geometry, also AI, machine learning, and published an article on that in January 2024 that uh claim or show how this program could solve Olympiad geometry problems. So, for people who don't know, the Olympiad problems refers to the International Mathematical Olympiad, one of the greatest challenges for the young students. Each year, more than 100 countries send teams of up to six students selected to very highly competitive national examinations to bring them to do one exam time for six questions. And these questions are incredibly difficult. These kids are so talented. So much so that since its inception in 1959, about half of the field's medalists have received, have been at this event and received medals. And some of the most famous ones have won a gold medal, like for instance, uh Tim Gowers. So these problems are really, really tough. So what they did was just look at the geometric problem, like the triangle and circle and all that. So that's not one of the IMO problems, just to show you how it works. They feed that information to the system, it says, oh, here is a triangle. Can you, if the two sides are the same, isosceles triangle, can you show that the two corresponding angles are also the same? Feed that in the machine, explore different possibilities through machine learning, and then have a part that's a symbolic manipulator of rules, and came up with the solution. Okay, that was the simple example, but then they apply that to the geometric problem of the Olympiad. Here is a 2015. I'm not gonna read it, but essentially you have to prove that two circles are tangent, and if if you just read it and you're not used to this problem, it's gonna make you head spin and all that. So these problems, they have a lot of these problems, and there's a systematic rule you apply and play around. That's what the kids know, but you really have to be very bright. And Alpha Geometry was perfectly fine doing this problem, so good actually, that on this type of problem, it's called a gold medal, just for this problem, probably fed to the computer, right? And that was just the start, because shortly after, in July 24, now the same type of program, a 28 out of 42, and that was already like a silver silver standard, right? And shortly after, a year later, just feeding directly the question just as they're given to the students within the same type limit, both Gemini Deep Think and OpenAI claimed the same result, gold gold medal for the for the system. Right? So people thought, well, that's pretty good. But again, if you're a little bit in denial state, you said, well, these problems are meant to be solvable. We give them to the student at finite time, and they have rules, and there's a big bank of that, of course, it's just manipulation, you know. This is this is not real mathematics. Just have a look at this problem and you tell me it's not mathematics. Then came the lover, spring of 2026. What can only be described as a brief but very intense love affair between machine and mathematician. To explain you the depth of it, I want to tell you a little bit what are famous unsolved problems. So in mathematics, uh unsolved problem is something that is in the literature that somebody asks at some point and that people are trying to solve. For instance, Fermas Last Theorem, and we'll talk about that, and other problems. And again, reading Tim Gowers, it would tell you that mathematical research consists in large measure in trying to solve unsolved problems. So, where do you find these unsolved problems? You find them in textbooks, but there are also people who are very good at finding interesting problems. And one of these people is Paul Erdosh, who is not with us anymore, but he was an extremely prolific writer with more than 1,500 mathematical papers, and also 1179 problems that he posed that are called the Erdosch problems. And Paul Erdosh was very well known, very eccentric. He was homeless for the part of his life, just going from university to university, just working with people. A wonderful story. And for instance, here is an iconic picture of Paul Erdosch with Terence Stau, Terence Tow at age 10, where he won his first bronze medal in the IMO at age 13, the youngest ever. To this day, he won the gold medal at the IMO, widely considered as one of the greatest mathematicians alive. So, one of Erdos' problems is a very famous one, and I want to explain it to you, because yes, we're in a crisis, but we should never let a good crisis go to waste, as they say. It's a good time to learn a bit of mathematics. And the problem is the following: how can we arrange endpoints in the plane so that as many as possible exactly one unit apart? I'll tell you a little bit more, I'll explain to you in detail. But it was widely known as the best known and the simplest to explain problem in combinatorial geometry. So let's start with, let's say, three points. Right? If I take three points, I can put them in a line. But you say, aha. If I move one of the points and make a triangle, now there are three pairs of points, three of these bars and not two. So by rearranging points in the plane, I can have more points that are one unit distance apart. Four, for instance, I can put in a square, that's a good idea, and then you can increase. And if you play, you can start playing around, and you find that the symmetric configuration is a good solution. And but then you find that if you go higher, you can change a little bit and you get a better result. But Erdosh's problem was really about the behavior of that for very large number of points. And EDOS not only posed the problem but also had a conjecture. And the conjecture was that if I place the point in a grid where the size is one over square root of m, where m is an integer, then I can really increase the number of points that are one unit distance apart, this green bar on the screen. And he computed exactly that limit, and the limit, the number u of points that are one distance apart, increased like n to the power one plus little o of one. That means a number that goes to zero at it goes to infinity. But it's bigger than n for any finite n. And his conjecture was this is the best possible arrangement that you can have. You can't really beat that. The limit always goes, the best you can do is n as n goes to infinity, or scales like n. And so in May 2026, people or OpenAI came with a different solution. A solution that looks completely different than all the construction that had been proposed, and is based on idea from algebraic number theory, and produced very, very large, highly structured point sets that beat the conjecture and that are n to the power one plus delta with a finite value of delta for very large n. And that was quite a shock to the community or people interested in that. And they post nice video interviewing uh interviewing mathematicians that say wonderful, this combination, it's very creative and all that, and that was great, right? But like sometimes love affairs don't last, and after May came the summer. And with the summer came the soldier seeking the bubble reputation. And in the summer, many different things happen. Uh, I'm gonna start with an example that that comes from slightly different fields, but I find very interesting and telling. In July uh 2026, it is July, uh a new article was uh posted by uh Dr. Shamnu Jin. And what he looked at is a problem called the Crusade conjecture. I'm not going to give you the detail, but it's a conjecture related to the behavior of complex matrices, and it's very important in numerical analysis. It tells you exactly uh it tells you or a polynomial of this matrix grow uh depending on the numerical branch of the matrix. And the conjecture was that there was an inequality with a factor two here, that was the best possible way to describe the problem. And it's been posed by Michel Cruset in 2004, and the best number there was up to uh up to the work in July, 2.414. And the conjecture was that you can lower that to two. And while I'm showing that without giving you detail, is because Dr. Chamnu Jin is not a mathematician, he's a neurosurgeon, surgeon in China, in Beijing, and he has interest in uh in geology, in his background, he had no formal education in mathematics, and he found the problem by looking at uh trying to learn about uh transcranial imaging, where this type of uh inequality comes in. He said, Ah, that's interesting. Why don't I ask ChatGPT? And here is to show you that's what the important part, the full prompt that he gave the machine. Right? So the prompt has two parts exactly the statement of the conjecture, the mathematical statement. You can take a screenshot and say, here's the mathematical conjecture. And below is the task asked by the computer to check the current statement. He was very open about it. Uh he gave the prompt, GitHub, and all that, the paper, everything. Uh wonderful, uh, wonderful attitude with respect to that. This is what we call a one-shot problem solving. Here is the problem. Try to do these things and give me the result. Right? Anybody can do that now. You can go into your old textbook and you can take a screenshot of that and put a prompt and just wait to see if your machine is going to solve it. So, what is the prompt? Here are a few lines from the prompt, but essentially it's what you would say to maybe graduate students go away and solve this problem and don't come back until you have the answer. Uh, he took it from a problem for OpenAI, the prompt that he was given. And the prompt says, give a rigorous standalone proof of the above math problem using your own knowledge, computation, and reasoning, return the proof as a full English file. When you don't write the full proof. Use agent throughout every candidate proof must be checked for gap. And this agent is interesting. An agent is not a little person, despite all the company trying to anthropomorphize this program. T is a little program. So since our LLM can take language and produce language, this can be fed to another LLM. So you can run essentially multiple LMN with slightly different personality. Different by personality, I mean different constraint. You know, you change, you tweak the different parameters in the underlying program, and you can make in track them together. You can have a program that's used as a boss to organize the work of the other ones and let them work together with different try. This program is gonna go do a certain task, this one is gonna do a certain task, come back, and if the work is good, blah, blah, blah, just like you would have in conversation. Since they take conversation, they can converse with each other. And this is one of the secret sources that that they use now. Return only when a complete affirmative proof has been found. Do not return a partial result, best effort summary or explanation of why the problem is difficult. You know, and do not look at previous literature either. Just do it on your own. Well, 16 hours on Chat GPT 5.6 sol, so purely one that we that is already odd-lift. This is an old one already, we we pass that, in 31 July, completed the proof. Michel Cruset looked at the proof, other people say, Yeah, it's interesting, it's correct, despite the fact that they had spent entire workshop in the past trying to make this proof, right? So imagine a little bit of things. Change. That's the end of July. Let's move to August. There is no greater result or best-known result in mathematics than Fermat's last theorem. So let me remind you that for positive integer, you can take 3 square plus 4 square equal 5 squared. They all are the possibilities. But if I change a square to a cube, then there is no integer satisfying that. And you can change to any power, and you can't find ABC integers so that AN plus B n equals Cn when n is bigger than 2. Beautiful theorem, Ferma's last theorem, that was proved in 1995 by uh Andrew Wiles and Richard Taylor to great fame. And this is one of the reasons why the Oxford Math Institute is in the Andrew Wiles building in recognition of this. A proof is accepted by the community. There is no problem with it, everybody agrees with it. But one of the quests was not so much the proof, but formalize the proof. Can we put it in the same language I told you about that would guarantee all the steps all the way to the end and give a certificate of the proof? It's a different type of mathematical challenge. We notice proof exists, but can we formalize it so that it's certified, every step is certified? And in the process, you learn a lot about the structure. And here is what uh Anthropic in this case did in August. And it's very interesting because they started, and it tells you the difficulty also that Fermer's last theorem by itself is at best an oddity. It's a nice little relation, it was never the point. What's fundamentally interesting with the Fermer's last theorem and the way it was proved is that mathematicians over centuries had to put, and Andrew Wiles at the end, had to put all these theories different and build new theories in order to prove that one theorem. So the interesting mathematical content is not the Fermer's last theorem by itself, it is the entire mathematical apparatus, all the things they discover, the bridge between different theories. But in order to formalize that, anthropic run the cloud for 11 days, right? In the process, this is the end of the of the proof. You know, these programs also have this what's called chains of thought, where chains of reasoning, but they tell you what's going on in their head, which is always very fascinating because he says, Oh, I'm doing this. No, I'm doing this, I'm grabbing that. That doesn't work, I'm gonna try that. And here is the last one. It said, historic moment for this campaign, modulo recheck. You might have to recheck, but you know, I'm done with that. And the the machine, I didn't, I wouldn't say realize that, but that was the last output. But the scale of that is just mind-blowing. Right? In the process, the machine proved 29,500 theorems and it wrote in his lean program 13 million lines of code. For the seven previous years, the entire community of mathematicians working on formalizing mathematics had written 2.5 million lines. It's a thousand times faster. The scale is just incredible. And for instance, Kevin Buzzert, very excellent mathematician at Imperial, had an entire uh EPSRC fellowship dedicated to formalizing or making progress towards formalizing Fermo's last theorem. And there was a press release in September. Now you say, oh, wonderful, and Kevin was very nice about it. It says interesting, it's a little messy, but we have a lot to learn about it. But things at the same time was going to turn a little bit more ugly. And we move into the canon mode. So, in order to tell you what's next in the story, I have to tell you about certain types of unsolved problem. They are famous unsolved problems, but they are very famous ones, the one that everybody wants to solve in mathematics, that would guarantee you instant fame if you manage to solve them. And that was started in 1900, the turn of the century at that time, by David Hilbert, great mathematician. And he says, here 23 problems to guide the entire field of mathematics that I consider an outstanding problem that needs to be solved. And they were really lighthouses for mathematics, guiding mathematics in different directions and building that. A lot of these have been solved now. For instance, well, for instance, one of them is the Riemann hypothesis that has not been solved yet, but I'll come back to that in a second. Problem eight. There was also problem 18, which is building space from polyhedra, which related to the Kepler conjecture on how to best uh do sphere packing in three or higher dimensions. So the way the people do at the grocery store in pyramid turns out to be the optimal way. And that was proved also by a computer uh was proved and verified by a massive computer verified proof in uh 98. This problem was so important that at the turn of the millennium, the Clay Mathematical Institute decided that to identify, there are many different eminent mathematicians coming, identify seven problems of great importance for mathematics that would play the same role. And to incentivize the community, it says they'll give one million if you solve it. Yes, one million dollars. Among them, you have the Riemann hypothesis. Again, come back to the Riemann in a second. You had the Poincaré conjecture that dated from 1904. It would probably be part of the uh Hilbert's problem, but it came only a few years later. That was solved by another very eccentric mathematician, Gregory Perelman, also gold medal at the IMA in 2002. He refused the price, he refused the money, refused the Fields medal, and why not, and stopped doing mathematics. It's an interesting story on its own. Another one was the so-called Navier-Stokes problem. And you might have heard because it was in the news. So let me tell you about Navier-Stokes, because in applied math, it's a problem or it's a type of equation we know very well. And the resolution of that was really a seismic event. So the Navier-Stokes equation is the equation that described the motion of fluid, very simply. But it's so universal that it managed to describe uh motion of fluid starting at 20 levels of atoms sliding together all the way to fluids that are in stars and things like that. It's really universal. It has its limit, but it's one of the great equations that we use every day in engineering, in physics, in various fields, to understand the motion of fluid and its consequence. Engineering, it's extremely important. Any pipe, any things, turbulence, it's very important. Aerodynamics on cars, on boats, on planes, all that. It's very well known. There is an entire field in science, compute, uh, computer fluid uh dynamics, which is only dedicated to numerically solving this equation. We know to solve them, we know plenty of cases. There's like gigantic literature. And the equations are fairly simple. It describes essentially F equals ma for a continuum. On the left side, you have the acceleration, u is the velocity, the derivative of u is the acceleration, but you move with the flow, so you have an extra transport term there. And on the right, you have the viscosity force, the forcing force f that turns out to be important in the problem, and the pressure that you need because you have an incompressible fluid, and that's the second equation tells you it's incompressible. It's a very good model in to first order, assuming that uh volume of a fluid cannot be further compressed. And so T's are not an easy equation to numerically solve. That's why there is an entire branch for that. And they're simple to state, but of course they can describe extremely complicated behavior. And here is an example, and that was given to me by Javi with a PhD student in Oxford. And I said, Can you give me an example? And so a single prominent chat GPT, and he produced this very nice uh related instability, which is instability of a heavy fluid on a lighter fluid. You know, like you have water and oil, and what happens as they start mixing. And you can see the problem with the complexity of the motion. So, view as a science problem, we know a lot about it, but view as a pure math problem, the question was whether the equation can support solutions that blow up to infinity. If you force it the right way, starting at rest, can you have a solution for which part of the fluid, the velocity actually goes to infinity? And there was a lot of discussion, that was the Millennium Price, formulated in a little bit more mathematical term. And uh people in fluid dynamics had already found numerically that there was the interesting solution that are vortices where the uh along the axis the velocity was increasing, maybe beyond Bob. And it was a lot of work in mathematics trying to solve that because it became a famous problem. Uh and mathematicians, especially uh in New York at the Kura Institute, were starting to see the possibility of finding a solution. They said, we can probably do this, and but there was a leak. And when the leak came out, that Tristan Bachmaster was close to getting uh to the solution, the big frontier lapses, we have to get there, and they started the machine on the problem. And what happened is on September 8th, a few days after Ferma's last theorem, everything is so fast, they came up with an entire solution of the problem. Uh they they said they have a lean certified code, but people arguing about the detail, about the certification in view of how they phrase it. But there is very little doubt. I think the consensus knows it's probably right. And within a few days, they produce an article of 165 pages with the full solution, which is essentially all functional analysis, very hard, very hard to read. Uh, and they started with solution close to T1 with the vortices and all that. Essentially solved the millennium problem the way it was asked. Again, the way it was done was incredible in terms of mathematics, things we've never seen before, right? The industrial scale. In this case, OpenAI used 10,000 concurrent agents, this little program running together, interacting, for 88 hours using 130 billion output tokens, whatever tokens are. I know it's it's cheap, but essentially it's a unit of information going to the machine. It's a lot. And it costs maybe the estimate as $22 million. And that's not even counting all the other Frontier lab, maybe Anthropic, maybe uh DeepMind or whatever, that have also tried and put money there. Never so much money has been put on a single math problem in so little time. It's completely unprecedented. And then the one formalized with GPT-6 Astra, which is actually one that is uh freely available. So this graph is interesting. They use the internal model, and what you see essentially on the y-axis is how good the model are, depending on how long you run it, in solving math problem. The one below, the line below, is the one for the commercial uh GPT, the 1.6 Astra, as good as the one, and their own internal model is much, much better at doing that. And so they realize on all these tests that all of a sudden they have something much, much, much better, and they could run it on other problems. And overnight, mathematics, pen and paper, the cheapest discipline became the most expensive one. Do you need a big computer like that in order to solve problems? And this is this is just the beginning because this model, the internal model from what the rumor is, is soon going to be released uh publicly. And so everybody will be able to use this type. Of course, we'll have to pay for it. Now, when you have such a seismic event, you know, in the ocean, sometimes it can create very large wave that creates a tsunami. And that's what more or less happened. What they did next, or at the same time, is they started running their internal model on a collection of 4,000 open problems in mathematics. And they just feed the machine and ask the machine to produce, approve, and write a paper. And there was already in the community rumors that they had found maybe another millennium problem, solved plenty, maybe 10, 20 problems that were big problems of mathematics. But nobody was ready for what came next. It was like a shock. In October 6, 2026, we knew the date from internal people were talking. I was checking at night, OpenAI released that. Because of the mess with the Navier stocks, the fact that maybe they use information but maybe not, or they use the leak, or the way they behave with mathematicians, there was already a lot of uh problem between the mathematical community and open AI. And in this case, rather than making a big fuss about it, they just quietly released to GitHub and let mathematicians take care of it. It's not on their web page, no press release or anything. This is what we found, this is now your problem. And what they found was not 20 or 100, was the solution of 372 open problems out of the 4,000. There's still plenty of mathematics to do, and 722 manuscripts in all areas of mathematics. That's what's shocking. The list is there, and there's number theory, there's algebra, there is uh geometry, there is probability, mathematical physics, differential geometry, partial difference, etc., etc., etc. It's believed now, people have said, the probably the single most important day in the history of mathematics. On 7 October, there were three withdrawals and 719 remained. The problem was that the computer used different conventions with sign coming from different theory, problem of notation, and that made problem with the proof, and that was spotted and removed. So we can go, we can be here for a long time talking about the 372, but one of the most important that was directly realized, and one of the only ones that was edited by human, the other ones were all machine written, was not the Riemann hypothesis, but quasi-Riemann. So what is quasi-Riemann and what is the Riemann hypothesis? Now it's time to tell you. So here is a very simple function, which is the sum on the integer of 1 over n to a power. So very simple. If I put s equals 2, then I have 1 plus 1 over 2 square, so the sum of the inverse square. And it was already known by Euler in 1737 that remarkably this gives pi square over 6. They are different demonstrations, beautiful result. But the important part, and why it's this function is so crucial in mathematics, is not for this type of result, is that you can rewrite this sum also as a product over the prime numbers. And prime numbers is the unit, the atoms of number theory. This is why it's they are so important. And Euler, this is the Euler product formula, also used by Gauss, realized that you can rewrite, for instance, zeta of 2, it's the zeta Riemann function as 1 over 1 minus 1 over 2 squared, blah, blah, blah. But here you see only the prime. So there is a deep connection between the zeta function and the prime number. So what is the Riemann hypothesis? Well, if I take S to be a complex number, what the function returns is also a complex number. That's the important thing. You need a little bit of math, so just uh trust me on that. The important thing is that there are special complex numbers S star for which the function returns zero exactly. And these are called the zeros of the zeta function. And why are the zero of the zeta function so important? Because they control the distribution of prime or the prime fluctuate around the average for very large number. A theory of prime is really trying to understand for a very large number what happens. And this hypothesis is central in all not only analytic number theory, but plenty of other fields like random matrix and all that, and it's widely considered as the holy grail. So, what is the holy grail is to prove that the zeros of this function, apart from the trivial ones that are on the axis, are in a narrow band all on S. Imagine a real part of S equals one-half. That's the only place where you can have zero. So that's the conjecture. The conjecture tells you the zeros can only be found on this one critical lines. And it's very important because plenty of other theorems in mathematics depend on this conjecture to be true. It is a called conditional result. It says if Riemann's conjecture is true, then I can prove this, this, and that. So it has a lot of foundational importance. It's not just like an oddity like Fermat's last theorem. And the consensus is that yes, it's probably true, right? But remember, unreasonable mathematicians, they want to be known as true. And as an example, people have computed up to the first 12 trillion zeros in that band, in that in that great, in that red band there, and they are all on the critical line. So it's not like there is that much doubt about it, but the real problem was to really to prove that. And so far, the best was that there was a narrow slip here in blue, a narrow strip where there was no zero, that was guaranteed. And what OpenAI came up with, Chat GPT, was a proof that there are two bands, independent of how you go, two straight bands between zero and one eight that you automatically have by symmetry, and seven, eight, and one, where there is no zero. And that was stunning for mathematicians. Like one mathematician says it's a result that I thought was complete science fiction. What about Phil's medalists? What do they have to say? Where Tim Gowers says it's a very, very big deal. The biggest step toward the Riemann hypothesis for a long time, of France Terence Tao says is the biggest advance on the subject in 50 years. And James Mannert, also Fils Medal, colleague at Oxford, says it's obviously the best line result for the past hundred years. If anybody, mathematician, would have come up with a result like that, again, it would be instant fame for them. But also very important because this result is actually enough for a lot of other theorems. You don't have to prove the full Riemann hypothesis. Let me tell you a more uh personal story. Uh I worked in the past of something called the Ericsson problem about the non-existence of solution in uh in in mechanics. And with my friend Arash Yavari in Georgia Tech, we managed to reduce the problem to a very large set of algebraic equations. And the question was to prove that the set has no solution. And on 20 October, I received what I was teaching, I received uh a text from uh from Arash. He says, Have you seen this? Have a look at this paper. I said, What paper? Well, the paper is called a local solution of Ericsson's problem, and actually cite a paper, that's how we found it in the first place, but actually prove the result that we've been trying to prove, not only with us, but with a graduate student that spent a year trying to reduce the problem. The graduate student, Christian Goodbreak, did very well and did plenty of other interesting things, but we certainly spent some time. And so Arash says, most likely this proof is AI generated. I said, how do we know? Well, the paper is 214 pages by an undergraduate student. So maybe it's very bright indeed, or maybe it's a one shot, as we had. So we don't know, right? And you can ask, or do I feel about it? Well, that's the main thing. How do we all feel about it? So after 6th of October, the world realized one thing is that mathematicians have also have feelings too. And the feelings, you can find them online because there was a website dedicated with that, proof and thrum, where people, a form of group therapy, posted their reaction to what's known now as the AI dump or the open AI dump in mathematics. And the reaction had diverse diverse level of uh different level of uh grief, if you want, but clearly it was shocking to everyone. Martin Bryson, who is the president of the Clay Institute, another colleague at Oxford, said, I never imagined the frontiers of mathematics could be moved so far in one day. The range of proof are just stunning, says another one. Solving problems I might not have expected to be resolved in my lifetime. This huge dump of paper is a tsunami. And the reaction on the on some people were really, really uh sad in a way. For instance, when mathematician, right, he was devastated, real stab in the back. He feels I'm standing among reeds, uh, runes, question that guided 46 years of my career, suddenly wiped out. Another one says, Colleagues are experiencing severe anxiety, depression, and loss of purpose. Every department in the world, people were talking about that over coffee. What does it mean? What is going to change, and all that. Hugo Dumenil Copin, other fields medalist, they've annihilated my field of research. He says, All the problems I've looked at have now gone. I really cried. And the disruption for students for early career researchers is huge. For instance, Terrestao says, I'm horrified by the many person years of patient and deliberately slow research effort being casually disrupted or destroyed. Clearly, the end of that love story. But otherwise, we're much more excited and hopeful. For instance, it feels like living in a mathematical wonderland, right one. People do math in lots of different ways for lots of different reasons. There'll be more ways to do math in the future, I believe. And so there was a sense of, okay, people realized this could be very interesting, right? But the main thing is that people also realized we have to radically expand a view of what a mathematician is and does. And so we came to the last part of the talk. It says, What is mathematics anyway? Right? We are in an existential crisis, the existence of the field of mathematics. It's a good time for reflection. What is mathematics? Is it really, as Tim Gower said, that our main job is to research and solve problems? Or as Sir Michael Lattia, also Fields Medal, who died a few years ago, I would have plenty of things to say, says it's a mistake to identify research in mathematics with the process of producing proof. Proof is the end product of a long interaction between creative imagination and critical reasoning. But if you need this in solved problems to get there as motivation, it's still a problem. So there are plenty of other aspects that people assess me. Maybe problem formulation is more important, or theory building. And so there is a deep reflection now about what is mathematics and what is the motivation behind mathematics. And James Mennard, for instance, says the primary goal of mathematics has always been about human understanding. So I'm more of an applied mathematician, even though I dabble sometimes in theorems and things like that. And people come and say, What's the status about applied mathematics? Well, you have the Old Testament and you have these newer testaments, which is a very short introduction to applied mathematics by me, where I look at what is applied mathematics. And in applied mathematics, we do not have unsolved problems. It's not something you find. What we have are grand challenges. And the grand challenges are climate, energy, trying to model, understand life, trying to model the brain, and other things like that that I've talked about in the past. It's very different. The big premium is not on solving a problem that's given to us 100 years ago, it's looking at a grand challenge and formulating a problem that can be formalized mathematically, for which you can make progress at the mathematical level. So problem formulation is really a premium. Theory building is also how do we generalize the theory that we have to include new effects that we see or combine different effects? Then, of course, you have a set of methods, and that's where we can really use this tool now. You have the ability to do so much more, so much faster, then all of a sudden you have a whole new landscape that opens. Because now you can think of the application and you can think that what are the big challenges and rethink that now that we have this amazing tool where we can do so much more, so much, so much resource, look at the data, understand, understand the problem, run numerics all at the same time. How should we restructure ourselves? And these are the discussion we have, in order to use mathematics to answer this question in a way fulfill the promises of mathematics. And it will allow us, it's already doing, finding this bridge to other scientific areas, but not only to our scientific areas, but also to bridge to pure mathematics. There are plenty of tools in pure mathematics that are very obscure. LLMs are not only not the only ones at writing good mathematical papers. There are also plenty of badly written mathematical papers, very obscure, very difficult to read. But now we have the ability to understand them, bring them, and using the power of mathematics, even of pure mathematics, to this problem. And so many of us in applied mathematics are very excited. We see that as a new golden age for mathematics in general, but certainly for applied mathematics. But what will it change for the field, the discipline? Well, I would say almost everything is going to change. Every single aspect of what we do and the way we do it is about to change. And it's very difficult because the pace is unsustainable, really difficult. For 2,000 years, we've been like this man here, you know, taking mathematics, pen and paper, thinking about the problem, which is a good way to think. It's a nice life and all that. And now this is what we have. Right? And I can tell you that jumping from one car to the next car is not conducive to do good mathematics. But now it's everything comes so fast, so difficult. It's like drinking from a fire hose, really difficult. And that's one of the problems. And another sad thing is that we went from the edge of trust to the edge of suspicion. All of a sudden, we can't trust anything we write as being the produce of product of human that we can have meaningful interaction with. In 2013, Yi Tang Zhang, who was a very completely unknown mathematician, came up with a beautiful proof of the twin prime conjecture, making a big breakthrough. If he came today with that, people would say, You didn't do that, you fed that on the computer. Come on, you know, how would we trust you? It's really, it's really, really sad. The other problem related to that is we had openness in mathematics. People exchange problem ideas, and now people are terrified. He says, I have students saying, I don't want to send this abstract before the paper is published because it's going to be fed into a computer and they're going to post the next day a paper on that using AI, right? It's terrible. So that that nice aspect of the community is about to change. People are not going to tell what their own problems are before they solve it. And once it's produced, who checks the proof and what do we do with that? Right? It's very difficult. And who gets the credit for the work? And for what part of the work do we get credit? These are all questions that people are discussing now, and that are all open. And also, what do we teach with these new tools? And that's an important question because I had people saying, coming to me, says, Oh my, you know, my kid is very good at mathematics, but should he still study mathematics? No, is it mathematics is that we hear in the news? Well, yes, unequivocally, more than ever, there's never been a better time to learn mathematics. There are new tools for learning, it is amazing tools. You want to understand a paper, you feed it in the machine and you have this Socratic exchange with the machine, trying to explain to you at your level what is in there. All of a sudden, the students can do that, we can do that, we can learn more. There are two great pleasures in mathematics: learning mathematics and doing research-finding stuff. And certainly the learning mathematics has all of a sudden expanded for all of us, not just students. But also there are this incredible new landscape to map. Now that we understand this, there are other things beyond it. We can make this giant leap. And there's going to be an emphasis on social math now, as much as we like the romantic idea of somebody, mathematician, working for 20 years in the attic and all that. Students, mathematicians, mostly are not like that. They like the exchange. And the exchange is going to be much more prominent now that we can't really rely on the paper. We'll have to exchange, we'll have to explain, do a better job at exchanging. And so that's great for students because typically students like that. But also, and I'll talk about that next year in a talk, we can really use mathematics for this big challenge, and students welcome that. They've been raised in an era of multiple crises, and going to mathematics, yes, they learned some aspect of mathematics, but plenty of them also like the idea that learning this could have an impact on studying, understanding, and solving this crisis. And finally, it's one of the greatest skill sets for our age in a in a in a landscape of information and all that where critical thinking is the most important skill, there is no better training than the mathematical training where logic is supreme. So I come to the end of the talk, but clearly it's just the beginning of this story, and I'm bound to have to answer the question is it the end of mathematics? since this was the title. Yes, all mathematics is death. But long live new mathematics. Thank you.