Mathematics has a good problem, and it is a big one. AI systems can now produce results at a pace that would have been hard to picture a few years ago, and the people trained to check those results need time, tools and places to compare notes. On October 15, 2026, in New York, the Simons Foundation announced a $15 million combined commitment with Anthropic and OpenAI to help mathematicians do exactly that.
The money is not a prize for a single breakthrough. It is support for the slower, essential work around the breakthroughs – reading, testing, questioning and deciding what counts as solid. Funding will be managed and awarded by the Simons Foundation, including through an open call for proposals. That detail matters. An open call invites researchers with different approaches to make their case, rather than limiting the effort to a small circle already close to the labs.
What the $15 Million Covers
OpenAI’s contribution was reported as $5 million, plus model access and computing resources for foundation-supported scientists. Anthropic made a similar $5 million commitment, including funding, compute and Claude access. Put together with the Simons Foundation’s own share, the combined commitment reaches $15 million in funding and resources.
Cash alone would not solve the practical problem. A mathematician who wants to test an AI-assisted result needs access to the models, enough compute to run serious checks, and colleagues who understand both the mathematics and the software. This package tries to supply all three at once. Access and compute sit alongside the grants, so supported scientists can work with the tools directly instead of reading about them from a distance.
David Spergel, president of the Simons Foundation, leads the organisation that will manage and award the funding. That gives the effort a clear home. The foundation can run the open call, select the programs and keep the focus on researchers and their needs, while the two AI companies contribute money, access and computing power.
Why Evaluation Needs Help Right Now
The scale explains the timing. On October 6, OpenAI released 722 manuscripts organised into 372 families of results, drawn from work on roughly 4,000 problems. Read that number again. Seven hundred and twenty-two manuscripts. Even a strong department could spend months working through a fraction of that output, checking steps, hunting for gaps and deciding which results deserve deeper study.
That is not a complaint about the manuscripts. It is a simple question of capacity. Traditional peer review was built for a world where papers arrived at a human pace. When results arrive in the hundreds, the community needs new ways to sort, test and discuss them, without lowering the standard for what gets accepted as correct.
The infrastructure story behind this shift is growing fast, too. Our recent coverage of Modal raising $750M as AI infrastructure demand climbs shows how much capital is flowing into the systems that run AI. The Simons effort asks a different question – who checks the work, and how well?
Workshops, Conferences and Shared Standards
The funding will support workshops, conferences and targeted programs. Researchers will use them to evaluate recent AI-assisted results, learn new tools, share perspectives and shape collective standards for AI use in mathematics. Those four jobs belong together. A workshop where people only swap tips on software would miss the harder questions about proof and trust. A conference that only debates standards, without hands-on time with the tools, would miss how the work actually gets done.
Shared standards sound dry until you need them. When is an AI-assisted step acceptable in a proof? What should an author disclose about how a result was found? How should a reviewer test a claim that came from a long chain of model suggestions? There are no shortcuts here, and individual labs answering alone will produce a patchwork. Collective standards, argued out in rooms full of working mathematicians, stand a better chance of holding up.
Beyond Pure Mathematics
The scope extends beyond pure mathematics into basic science. That widening makes sense. The same pattern – models producing candidate results faster than experts can verify them – is showing up across research fields, and the habits built in mathematics tend to travel. Mathematicians are strict about definitions and proof for a reason. If they can build workable norms for AI-assisted work, other sciences will have a tested example to adapt.
Training data sits close to this story as well. As Labelbox raised $100M plus to train AI agents by practice, the broader lesson is that quality improves when experts stay in the loop. Mathematics is applying the same idea from the other direction – keep experts in charge of judgement, and give them better support.
Practical Takeaways for Researchers
First, watch for the open call. Funding managed and awarded by the Simons Foundation, including through that call, means researchers outside the two companies can apply to take part. Second, treat model access and compute as part of the grant, not extras. A proposal that pairs a clear verification question with a real need for Claude access, OpenAI model access or computing resources fits the shape of this commitment.
Third, think in groups. Workshops and conferences are built into the plan, so collaborative proposals – teams that can test results together and report back what they learned – match the intent. A lone researcher can do fine work, but collective standards need collectives.
And fourth, do not wait for perfect norms before starting. The programs exist so researchers can evaluate recent results and shape standards while the work is fresh. Standards written after the habits have hardened rarely change much.
There is a quiet optimism in this announcement. Nobody here is pretending AI output should be taken on faith, and nobody is suggesting mathematicians step aside. The bet is the opposite – give the experts funding, access, compute and rooms to argue in, and let them set the terms. $15 million will not settle every question AI raises for mathematics. It should, though, buy the field something it badly needs right now: time to check the work properly, together.
Be First to Comment