A CERN for AI-assisted science?

Mathematician Dimitris Koukoulopoulos called for publicly funded frontier AI systems for academic research in a guest post responding to OpenAI's announcement of a solution to the Navier–Stokes Millennium Prize problem and the controversy over how it was obtained and released. Koukoulopoulos argued that concentrating frontier AI access in a handful of private companies threatens research confidentiality, unpublished results, and scientific incentives, citing the case of Julia Stadlmann, whose first improvement in more than a decade on the record for bounded gaps between primes was surpassed by AI-assisted efforts within days after two years of work. He proposed free baseline access to conversational and agentic interfaces backed by substantial public compute, with projects needing more resources applying for multi-agent systems and large-scale compute.

This is a guest post by Dimitris Koukoulopoulos https://dms.umontreal.ca/~koukoulo/ . This blog post was initially written in a different file format and converted using AI. — T. At the 2026 World Cup semifinal, England led Argentina 1-0 with less than twenty minutes left. And then the momentum completely changed: England switched to a much more defensive formation, withdrawing attacking players and retreating deeper. The Argentinians started attacking in waves. They equalized in the 85th minute and scored the winner in stoppage time. England’s defeat offers a valuable lesson: when trying to avoid the worst outcome, fear and passivity are not your friends. The recent announcement by OpenAI of a solution to the Navier–Stokes Millennium Prize problem https://openai.com/index/navier-stokes-solution/ , and the controversy surrounding the way it was obtained and released https://mathandai.org/ , should be a wake-up call for the scientific community. We need to come together quickly and develop an institutional response to this new reality. Complacency and defeatism are not an appropriate response, or else we risk turning the worst scenarios of AI disruption to science and society into self-fulfilling prophecies. One thing seems clear: AI is here to stay. These tools already have spectacular and potentially transformative capabilities, and we need to learn how to incorporate them fully into scientific research. But if we want to do so in a way that is truly beneficial to science and society, I think we urgently need publicly funded frontier AI systems for academic research. These should include simple conversational and agentic interfaces with substantial public compute behind them. Researchers would have free baseline access to them, while projects requiring more substantial resources would apply for access to multi-agent systems and large-scale compute. There are several problems with allowing access to frontier AI to remain concentrated in a handful of private companies. First, scientific research can involve sensitive, proprietary or even classified data. Researchers need environments with strong confidentiality protections and clear accountability in the event of data breaches. Second, unpublished research itself is sensitive. Prompts, uploaded documents and interactions with AI agents can contain new ideas and partial results. Researchers need strong guarantees that this material cannot become available to systems that might subsequently reproduce or build upon it. The recent controversies surrounding Navier–Stokes and non-sofic groups https://terrytao.wordpress.com/2026/09/11/on-the-existence-of-non-sofic-groups/ illustrate how even the possibility of such reuse can seriously undermine researchers’ trust in these systems. Third, the incentives of private companies are not the incentives of the scientific community. Competition encourages rapid demonstrations of new capabilities and spectacular results. Speed is valuable in science, but so are proper verification and attribution, careful exposition and, crucially, human understanding. For example, in mathematics, which is my field of research, a theorem is most useful to the community not simply when it has been proved, but when its ideas have been digested and incorporated into our collective knowledge. Speed can also be harmful, particularly for young scientists who are just establishing themselves. The recent experience of Julia Stadlmann https://www.sciencenews.org/article/human-math-record-twin-primes-openai is instructive: she worked for two years to produce the first improvement in more than a decade in the record on bounded gaps between primes, and then her result was surpassed by AI-assisted efforts within days. The point is not that stronger results should be withheld. It is that we risk moving to a system where years of early-career work can be eclipsed almost instantaneously by organizations with much greater resources. Young researchers are the future of science; it is our duty as a society to nurture them. Public access to frontier AI would reduce the enormous asymmetry in the research tools available to scientists, particularly to early-career researchers. This brings me to my fourth objection: access to the strongest AI systems is currently highly unequal. The Navier–Stokes result, for example, was produced by an internal OpenAI model that the company describes as significantly more capable than its publicly available frontier model. Several other recent mathematical advances https://openai.com/index/ten-advances-in-mathematics/ were likewise obtained using unreleased models. This risks creating a two-tier scientific system: a small group of researchers selected by, collaborating with, or employed by AI companies may have access to research capabilities unavailable to everyone else, regardless of what subscription they are willing to pay for. Public AI cannot prevent private companies from pursuing scientific results. But it can ensure that the scientific community has an independent, powerful alternative. To address these issues, we need to build a CERN for AI-assisted science. This analogy is not new: the European Commission has itself used it https://digital-strategy.ec.europa.eu/en/news/eu-launches-investai-initiative-mobilise-eu200-billion-investment-artificial-intelligence in describing its AI infrastructure plans. What I am proposing is a more specific version of that idea, centred on AI as public infrastructure for scientific research. The timing may be unusually favourable. The EU and Canada are rapidly deepening their strategic relationship. As I write this, European Commission President Ursula von der Leyen has proposed that Canada become the EU’s first-ever “associate member” https://www.reuters.com/world/eu-opens-door-canada-become-first-associate-member-eu-commission-president-says-2026-09-16/ , while Mark Carney is in Strasbourg https://www.pm.gc.ca/en/news/news-releases/2026/09/12/prime-minister-carney-strengthen-canadas-ties-european-partners-visit to address the European Parliament. This constitutes a remarkable step toward what Carney has called a “unique alliance” https://www.reuters.com/world/americas/carney-pushes-idea-making-canada-associate-member-eu-wsj-reports-2026-09-13/ between Canada and the EU. In addition, both sides are making major public investments in AI and computing infrastructure. Indeed, Canada has committed C$2 billion over five years to sovereign AI compute https://ised-isde.canada.ca/site/ised/en/canadian-sovereign-ai-compute-strategy , including up to C$1 billion for public supercomputing infrastructure . The EU is developing RAISE https://research-and-innovation.ec.europa.eu/strategy/strategy-research-and-innovation/our-digital-future/european-ai-science-strategy/raise-resource-ai-science-europe en and a network of AI Factories https://digital-strategy.ec.europa.eu/en/policies/ai-factories to support AI research and innovation. It has also launched a call for up to seven AI Gigafactories https://digital-strategy.ec.europa.eu/en/news/eu-launches-ai-gigafactories-call-boost-europes-computing-capacity-and-unlock-more-eu30-billion , backed by up to €10 billion in public funding, and expected to unlock at least €20 billion in additional private investment. Even more importantly, Canada and the EU have explicitly agreed https://www.canada.ca/en/innovation-science-economic-development/news/2025/12/joint-statement-of-the-first-meeting-of-the-canada-european-union-digital-partnership-council.html to explore cooperation on fundamental AI research, agentic systems for scientific discovery, access to advanced AI infrastructure, and the co-development of advanced AI models for the public good. Such an initiative should be open to other research partners willing to contribute resources and subscribe to common standards of scientific governance, confidentiality, access and accountability. A CERN for AI-assisted science cannot be merely another supercomputer centre. The user-facing layer is essential. Every researcher should have access to a simple chatbox to directly access frontier AI models without having to become an expert in model deployment, cloud infrastructure or GPU computing. Such a system would be a sophisticated research assistant, helping academics explore hypotheses and conjectures, test their research strategies, perform bibliographic searches, and many other important tasks. More ambitious projects could decompose a difficult research programme into many interacting subproblems and use coordinated AI agents to explore them semi-autonomously, with researchers directing the overall strategy and interpreting the results. Behind this simple interface would sit the expensive infrastructure: publicly controlled frontier models, large-scale compute and storage, and the teams needed to maintain and continually improve the system. In short, let’s give every researcher free baseline access to a world-class conversational and agentic AI research environment backed by public frontier AI and compute. Ordinary use should be effortless. Projects requiring exceptional resources could apply competitively for large compute allocations and sophisticated multi-agent systems, just as researchers apply for access to other major scientific facilities. Canada already has experience with this kind of mechanism: the Digital Research Alliance of Canada https://www.alliancecan.ca/en allocates advanced computing resources through a peer-reviewed Resource Allocation Competition. My motivation is not primarily to save academics the cost of AI subscriptions. I am much more concerned about the long-term independence of research. With AI becoming central to science, the fundamental research tools should not be controlled almost entirely by a handful of private companies whose objectives will not necessarily remain aligned with those of the scientific community. There are obviously many questions to consider: whether publicly funded models could realistically remain near the frontier; whether we should train models from scratch or combine public infrastructure with commercial and open-weight models; how such an institution should be governed; what scale of funding would be required; how access to very large compute resources should be allocated; and how to prevent the public infrastructure itself from becoming bureaucratic, technologically stagnant or captured by particular national or commercial interests. We need to engage in a public debate among scientists, governments and funding agencies about the feasibility of this project. The choice is not between using AI and rejecting it: AI is already becoming part of scientific practice. The question is whether the infrastructure on which future research depends will be treated as a public good, or whether we will allow science to become structurally dependent on a small number of private companies. That is not a choice we should make by default or through inaction. Just ask the English fans https://koukoulopoulos.gr/apaitoyme-na-leitoyrgisoyn-ptolem-v-ag-dimitrios/ . Acknowledgements: I would like to thank Patrick Allen, Eyal Goren, Andrew Granville, Paris Koukoulopoulos, Matilde Lalín, Daniel Litt, James Maynard, Carlo Pagano and Oscar Alberto Quijano Xacur for their helpful comments while preparing this text. I used chatGPT to sharpen the writing and perform literature research.