{"slug": "artificial-intelligence-in-research", "title": "Artificial Intelligence in Research", "summary": "A PhD candidate's September 2026 thesis preface recounts that an AI, or \"a swarm of them,\" solved the Navier–Stokes problem, one of the Millennium Prize Problems, marking a dramatic shift in how mathematical research is conducted. The author describes LLMs progressing from being \"useless at maths\" to performing standard calculations faster than him, checking his reasoning, and suggesting new ideas by March 2026, after spending roughly a year on the work in Chapter 5 of the thesis. The account frames 2026 as a period in which coding and mathematics are being automated, following ChatGPT's first release in 2022.", "body_md": "*Adapted from the preface of my PhD thesis, September 2026. Chapter references refer to the thesis.*\n\n## Past\n\nFor a long time, the term *artificial intelligence* seemed way too strong for the kinds of models I worked with during my studies and, more recently, in my research. A ResNet or U-Net architecture capable of classifying cats and dogs has, of course, a very limited form of intelligence.\n\nI remember watching YouTube videos back in 2015 about neural networks learning to play Super Mario World ([the video](https://www.youtube.com/watch?v=qv6UVOQ0F44)). I felt I was already late to the party: these models were already driving massive progress in image processing and computer vision as a whole.\n\nI also remember watching a documentary series about deep learning with my older sister five years later, in 2020, during the COVID-19 pandemic ([the series](https://www.youtube.com/watch?v=UwsrzCVZAb8)). I had only just started my computer science degree, but I had already decided I wanted to work in this incredible new field, convinced, probably a bit naively, that CNNs would help tremendously with real-world challenges. I remember my older sister asking: “*But what’s the end goal of this technology?*” to what I answered something in the line of “*I don’t know, building a sort of god that can solve all our problems, probably.*”\n\nLittle did I know that the real AI revolution would start in 2022, with the first release of ChatGPT. Oddly enough, I did not pay too much attention to it at first: I loved computer vision, but was not that interested in natural language processing. It was only in the summer of 2023 that I took out a paid subscription to an AI service, using it mostly for small bits of Python code at first.\n\nThis week, in September 2026, as I wrote the final words of my PhD thesis, an AI (or rather, a swarm of them) solved the Navier–Stokes problem, one of the Millennium Prize Problems. The name *artificial intelligence* seems very fitting now.\n\n## Present\n\nThis is a very odd time for research. I could lament indefinitely over the fact that the two activities I loved most, coding and maths, are being automated. But now is the time to pause and assess the situation honestly.\n\n### How my work changed\n\nMy day-to-day work has changed dramatically in 2026, and as I write this, the year isn’t even over. Here is an anecdote that sums up how far we’ve come in such a short time.\n\nI spent about a year on the work presented in Chapter 5. First, I read in depth about Bregman methods in optimization, the EM algorithm, and the link between the two. Then I formulated my algorithm within this framework and tried to prove the main property needed for convergence. When I realized that some important convexity assumptions did not hold, I searched the literature for ways around this and tried to build a proof using tools for non-convex problems. I spent several months on this before realizing it was a dead end and turning to a different body of literature for another line of analysis. The result of this year of work is Chapter 5, which I’m very proud of, even if the mathematical results fall well short of what I initially envisioned.\n\nOver those months, asking LLMs about my mathematical problems became a habit, a sort of personal benchmark. I would regularly ask them to prove, in a single attempt, the main result I was aiming for: convergence of the full sequence of iterates of my algorithm in the convex setting. They went from being useless at maths to doing fairly standard calculations better and faster than me (think of computing a gradient or Hessian), then to checking my reasoning for errors and sometimes even suggesting new ideas. I particularly liked that they always tried to offer some kind of reasoning, even when that reasoning was obviously flawed: it gave me something to react to and helped me get going when I didn’t know where to start. By March 2026, I was already very satisfied with how they helped me with my research.\n\nTo my surprise, in June 2026, a newly released model actually found a decent proof of convergence for my algorithm, with no errors that I could see. The proof was not particularly elegant. It just used the tools I had already tried, but in the correct order, and it worked. It took the model three minutes to work from my notes, write the proof in LaTeX, and compile the PDF. The model was working from notes that had taken me months to develop. Still, seeing it finish a proof I had struggled with for so long in just three minutes was depressing. All it took was a few cents’ worth of inference from an AI system anyone can access.\n\nSimilarly, coding has changed drastically over the past six months. I had been using LLMs for coding for almost two years. They helped quite a lot, though their code was far from perfect and the amount of context they could handle was always a limitation. Beyond chatting with LLMs, AI-assisted autocomplete was probably the feature that felt most natural to me, like an extension of my own arm.\n\nSince the arrival of coding agents, coding has become entirely different. I write almost no code: I tell my agents what I want, and they implement it. If I don’t like their style, I point them to my preferred style guide. If I have a bug, they fix it for me. I don’t recall being stuck on a bug for more than half an hour in the past few months, and I wasn’t even really stuck: the agents were. I spend most of my time reviewing their code and telling them which solution I’d rather see implemented. It is a very different job, closer to maintaining my own code or being a junior manager in a company.\n\nI estimate that maths and coding accounted for around two-thirds of my working hours, with the rest spent on reading, teaching and administrative tasks. Both maths and coding now largely involve checking and directing the work of agents.\n\n### Misaligned incentives\n\nThe rise of AI has accelerated the rot already spreading through research. Many issues in science and academia stem from the gap between what advances a researcher’s career and what makes for quality research. Of course, what counts as “quality research” is subjective. It is easier to describe the kind of work I am criticizing: superficial papers that offer little beyond what already exists, and give us little reason to care.\n\nAcademic careers depend on both the quantity and the quality of research output, but these are judged through imperfect proxies. Quantity usually means the number of articles published over a given period, for instance during one’s PhD. Quality is inferred from the prestige and visibility of the conferences or journals one publishes in, and from citation counts: how often other researchers refer to one’s work.\n\nThese metrics matter both to early-career researchers trying to secure a tenured position and to those who already have one. To recruit PhD students and postdocs, principal investigators need grants, such as those from the ANR in France or the ERC in Europe, whose allocation also depends in part on the same metrics.\n\nOne way to game this system is to produce many low-stakes, incremental papers, submit them to prestigious venues, and rely on the variability of peer review to get a few accepted. This boosts publication counts while also satisfying one of the proxies for quality.\n\nWhen it comes to citations, reviews and surveys are particularly rewarding. These papers synthesize a subfield or compare a range of methods, helping readers make sense of existing research. They are widely cited because they bring many works together in one place, making them convenient references for the introductions of other papers. Written by experts, they can be very useful, but they are also being churned out by early-career researchers and PhD students looking to boost their citation counts. So much so that review papers are now banned from arXiv, the largest global preprint server.\n\nAnother way to attract citations is to actively promote one’s work. Networking and marketing have invaded research: publishing an article now calls for a LinkedIn post to increase its visibility.\n\nI have not been immune to these incentives myself. Chapter 6 comes from my first-ever research article, written in early 2024: a very preliminary and incremental attempt to combine two algorithms and get some results.\n\n### The limits of productivism\n\nAIs can now generate these kinds of low-stakes papers almost autonomously, complete with a LinkedIn post and a website. Conferences and journals are being flooded with hastily produced contributions. Between the 2025 and 2026 editions, submissions to ICLR increased by approximately 68% to reach around 19 000, while those to ICML roughly doubled. This year, more than 60 000 papers have been submitted for ICLR 2027. The exponential trend is plain to see in the annual submission figures.\n\nResearch relies on qualified researchers volunteering their time to review each other’s work. While the number of papers has exploded, the number of reviewers has remained stable, putting enormous pressure on the system. Journals struggle to find available, qualified reviewers and increasingly turn to students, early-career scientists, or researchers in adjacent fields. Conferences face the same shortage and often require authors to review other submissions. In practice, this leaves much of the reviewing to students working outside their own subfield. Review quality has suffered, and decisions can feel arbitrary.\n\nDrastic measures should have been taken to limit the number of papers. Publication and citation counts should never have become measures of a researcher’s work. Applying a capitalist, productivist logic to research has encouraged us to equate producing more papers with contributing more to science. We now collectively produce more papers than we can digest, many of them hardly worth reading, with eye-catching conclusions drawn from insignificant experiments.\n\nAs in the wider economy, groups that accumulate capital through grants can recruit more students, who in turn produce more papers. Research is an industry. As LLMs take on more of the research work, access to computing power and paid subscriptions to the best AI models gives those who can afford them a significant edge over their competitors.\n\n## Foreseeable future\n\nNow for the little guessing game of predicting the future. I have realized that, unlike in other areas, I am completely unable to predict how AI capabilities will evolve, even over the next six months. Instead, I will try to anticipate the near-term effects of current systems.\n\n### The publication system\n\nThe cost of producing papers that would have been considered decent just a few months ago will fall even further. Heck, even good papers will become cheap to produce, and, dare I say it, even amazing ones, such as the solution to the Navier–Stokes problem. I’m sure solutions to many more famous problems will follow. The only work that won’t become cheaper is experimental work, but it accounts for only a fraction of scientific papers, and not the fraction I am personally interested in. Publication venues, whether conferences or journals, will be completely overwhelmed by the sheer volume of submissions. The incentives encourage everyone to try their luck at publishing large numbers of decent AI-assisted papers.\n\nThe only rational response, if we want to save our current publication system, is to use AI reviewers to spot flaws in papers and reject them automatically. TMLR, a well-known machine learning journal, is already experimenting with such systems. Many authors, reviewers and editors are now calling for a round of AI reviews for all submissions. The next step will, of course, be for those same people to try to game the AI reviewers by making their papers look stronger, though I’m not yet sure what form this will take.\n\nObviously, I don’t want to take part in this ridiculous game. Researchers who advocate for AI reviewers will end up like grains of wheat between two AI millstones. The prestige attached to publishing in top venues will decline even further, until the choice of venue means almost nothing. This is already partly the case: many acknowledge the randomness of the review system.\n\nStill, I do not think the current system will collapse abruptly. This quote sums up my own point of view almost perfectly:\n\n “The conference / peer-review model is dead and its corpse is moving only because so many moving things are attached to it.”\n\n— François Fleuret\n\nMany are already coming to the publication system’s defense, trying to patch it up and keep it alive. It will be a long and agonizing death.\n\n### Research\n\nThe publication system dictates how we assess research and the people who do it. If publishing a paper no longer tells us much about the work behind it, how do we assess that work? The publication system is central to research, and its death will have a huge impact on how we organize it. How do we judge the value of a PhD thesis? How do we assess a PhD candidate, or someone applying for a research position or a grant?\n\nBeyond that, the day-to-day work of researchers is already very different from what it was when I started my PhD. I have no doubt it will continue to change dramatically. In the immediate future, I expect many researchers to spend their days managing agents that run experiments and write papers to feed the zombie publication system.\n\nFurther down the road, as theoretical analysis and simulations become cheaper, I expect more scientists to turn towards experimental work. Collecting data and carrying out physical experiments seem harder to automate than writing code for simulations. For researchers who already work between theory and applications, this could mean closer collaboration with colleagues who conduct experiments, and more time spent on the practical questions that arise from their data.\n\nFor those of us who enjoy theoretical work far more than experimental work, as I do, I have little hope. Some may spend their time deciphering proofs produced by AI systems. Others may set large swarms of agents to work on the hardest problem they know and devote all their energy to finding a solution. At some point, we will simply be unable to keep up with these systems, and the value we contribute will diminish over time. Maybe our value will lie in explaining to other humans what AI systems have found, but that too can be automated.\n\n### Society\n\nMany of us researchers work in public institutions and enjoy some protection, at least in Europe. Even as a student, I do not think finding a postdoc in my field will be an issue for a few more years, perhaps even a decade. I have no doubt that the rest of society will soon feel the impact of AI, probably more severely than we do.\n\nMuch of office work is done entirely on a computer. I believe that, as of 2026, agents are beginning to be able to do this work, and that by 2027 they will probably be fully capable of it. Once agents can do this work reliably, I expect companies to start cutting jobs, not out of inherent evilness, but for purely economic reasons. Engineers, especially those working in fields closely related to computer science, will be among the first affected. Junior hiring will freeze first, as it is easier to stop hiring than to lay people off. It will become even harder for young people to gain the experience required for most entry-level jobs.\n\nI expect this to make youth unemployment much worse. With Europe’s aging population, this might also deepen inequalities between those who inherit wealth and those who need a wage to live. Universities will feel the consequences too. Why spend years studying for a profession that no longer hires? If job prospects disappear, the classrooms will inevitably start to empty. Fewer students will go on to do PhDs, and the number of researchers might decrease for the first time. Research itself could start to shrink, although I expect this to happen much later than in industry. Mass youth unemployment will lead to unrest. With this unrest added to the geopolitical tensions we already face, we are in for a troubled period.\n\n## What should we do?\n\nI won’t pretend to know how to solve all the problems I have described. Instead, I want to describe the kind of researcher I would like to become and the steps I plan to take to keep having fun and enjoying my work. This is my personal answer to a difficult question, and I believe everyone should seek their own.\n\n### Acquiring knowledge\n\nI want to cultivate both depth and breadth of knowledge through reading and working through problems. I want to understand where my field stands and, more importantly, how it connects to other fields, so that ideas can cross-pollinate between them and lead to progress. This takes years, and it seems clear to me that I should spend most of my time reading and understanding new topics. This is probably the real bottleneck to progress. For a long time, we have favored depth of knowledge in niche and narrow subfields. It is now easier than ever to achieve this depth of knowledge, let us use this to increase breadth. This would also help me work more effectively with AI systems and approach problems from new perspectives, something they still seem to struggle with.\n\nI will of course try to tackle harder problems, not by brute force but by finding elegant ways to combine existing objects and tools. Even without AI, it is possible to do “dirty” applied maths, with unverifiable assumptions, computational tricks and useless results. I have always tried to avoid this ugliness, and I believe doing so has become even more necessary now. The good thing is that I have never particularly liked writing technical proofs. What I enjoy is the moment when different ideas simply click together and the result we were aiming for becomes obvious.\n\n### Sharing knowledge\n\nI want to share what I learn as widely as possible, through writing, conversations with colleagues, videos and blog posts. Open science is of course a minimum requirement for me: both my code and my ideas should be openly available. I have always loved teaching and sharing knowledge, and I’ll keep doing just that. Videos are great, but regular classes and talks are also fantastic ways to teach. Many topics in maths, even relatively accessible ones, are barely covered online, whether in videos or blog posts. I feel I can contribute here by making them more accessible to colleagues and students. I would start by bringing drawings and prose back into maths papers. I also want to make it a priority to help younger colleagues find their way into my field, from the basics to more advanced topics. Through teaching, I hope to help them understand what took me years to learn, in a fraction of the time. I would also like to spend more time refining existing proofs and making them accessible to a wider audience.\n\nI don’t want to play the petty game of the dying publication system. Instead, I would like to find a better way to keep others up to date with my research. I can keep publishing preprints, but I doubt this is the best way to share knowledge. I imagine something closer to the way code is shared and maintained through version control. Maybe a permanent blog organized by topic, where I explain what I have learned and, hopefully, what I discover. The equivalent of a paper could be a package release, complete with the code for the experiments and their results. I would also like reviews to be open, much like comments on a pull request. Anyone could offer their opinion in good faith, with respect and kindness. I hope this would encourage constructive feedback, rather than the rushed, careless and often useless reviews currently protected by anonymity.\n\n### Collaboration\n\nScience, with all its flaws and strengths, is above all a way for humans to work together to solve problems and understand the world. I have seen firsthand how little researchers collaborate, even within the same country, city, lab or research team. Duplicated work is a waste of time, energy and public money. So is repeatedly building benchmarks to compare a new method with all the latest methods in the literature. Many of these comparisons are not conducted properly: competing methods receive less time and care than the proposed method. Every single paper claims to advance the state of the art, which is neither possible nor desirable. Optimizing for arbitrary metrics is a waste of time. I would rather contribute to shared public benchmarks that make these comparisons as quick and cheap as possible.\n\nI also want to welcome others who explore and develop my ideas. Realistically, I do not have enough time to pursue every idea I have. I’m very confident in my ability to leave problems unsolved and ideas undeveloped, so I might as well share them with the community. One possibility would be to keep a public list of research ideas, marking those I’m actively working on to help avoid duplicated effort. I also think this would make it easier for people to reach out when they see that we have similar ideas.\n\nCompetition is not the right model for research. It has always been, and should remain, a collaborative activity. Large conferences do little to encourage this kind of collaboration and mostly serve as networking events. I find week-long workshops with a few dozen people much more interesting: they leave room for real discussions and real work, beyond small talk. At the same time, trying to collaborate with people who do not share my commitment to open science feels wasteful. I have little tolerance for treating research as an individual endeavor whose purpose is to advance one’s career. I believe most obstacles to collaboration come down to differences in values and in how we think research should be done, rather than technical details.\n\nI do believe that this AI revolution can help us bring even the most abstract ideas into applied fields and, ultimately, put them to use in the real world. This transfer used to take decades, even centuries. Current AI systems can help us shorten this process tremendously by accelerating every step: working out the maths, explaining it clearly to colleagues, writing code to test the resulting methods, and finally applying them to real data and problems. I will try my best to keep real-world problems at the centre of my work, because that is ultimately what has always mattered, both to me and to the general population, which I’m supposed to serve. In particular, I’m still confident that developing new theories to analyze algorithms can lead to better methods and, in turn, better practical solutions.", "url": "https://wpnews.pro/news/artificial-intelligence-in-research", "canonical_source": "https://tmodrzyk.github.io/blog/2026/artificial-intelligence-in-research/", "published_at": "2026-09-22 22:57:02+00:00", "updated_at": "2026-09-22 23:23:48.042742+00:00", "lang": "en", "topics": ["artificial-intelligence", "large-language-models", "ai-research", "ai-safety"], "entities": ["ChatGPT", "Navier–Stokes problem", "Millennium Prize Problems", "ResNet", "U-Net", "Super Mario World", "Bregman methods", "EM algorithm"], "alternates": {"html": "https://wpnews.pro/news/artificial-intelligence-in-research", "markdown": "https://wpnews.pro/news/artificial-intelligence-in-research.md", "text": "https://wpnews.pro/news/artificial-intelligence-in-research.txt", "jsonld": "https://wpnews.pro/news/artificial-intelligence-in-research.jsonld"}}