Simons launches $15 million effort to help mathematicians assess AI

The Simons Foundation announced a $15 million mathematics initiative with Anthropic and OpenAI on October 15, 2026. The commitment combines funding and resources for researchers evaluating and using AI.

In an update dated October 15, OpenAI specified a $5 million contribution to the initiative, plus access to models and computing resources. The contribution is part of the overall commitment.

The practical issue is how mathematical claims become usable knowledge. A result needs an argument that other researchers can inspect, a clear account of its assumptions and enough explanation for someone else to build on it. Paying for that work changes who has time to do it.

How support will be awarded

Simons will manage and award the funding, including through an open call. Support will cover workshops, evaluation of recent results, tool training, community discussion and standards. Anthropic and OpenAI will also supply tools and computing resources to researchers supported by the foundation, extending beyond mathematics into basic science.

The announcement does not itemize each partner’s share or separate the cash total from the value of resources.

OpenAI says the money will support workshops, conferences and special programs. Its October 6 mathematics announcement had already described plans to fund work around understanding AI results. The named contribution and Simons arrangement appear in the later update. Keeping those dates separate matters when tracing what was actually committed.

Application details are still pending

The announcement does not specify eligibility, award sizes, deadlines, or the amount and duration of model access.

When ByteForward checked the Simons funding directory on October 15, there was no clearly identified application listing for this initiative.

For prospective organizers, a useful preparatory step is to identify a specific review problem and the expertise needed to address it. A proposal could explain which result needs attention, what participants would produce and how others could inspect the outcome. Those are practical planning suggestions, not published selection criteria.

Why funding mathematical understanding matters

The Advisory Group on Mathematics and Artificial Intelligence’s September 29 recommendations argue that labs should help finance human understanding of their mathematical output. They propose that established nonprofit institutions decide which work receives support, with mathematicians directing the intellectual agenda. Possible activities include focused workshops, working groups and detailed explanations of difficult results.

The same statement warns that accepting support does not endorse a lab’s practices. It also asks labs to improve attribution and presentation before release, rather than leaving all the repair work to mathematicians. Funding can support examination of a claim while the claim itself remains under review.

Our assessment is that the program will be most useful if its outputs remain available beyond the people in the room. A workshop can leave behind an explanation, a corrected statement or a reusable proof component. Those outcomes let the next researcher start from a stronger position.

Earlier work shows what a workshop can produce

A concrete precedent comes from Simons’ account of a June 2025 Lean workshop. The two week event brought 57 researchers together in New York. Participants formalized undergraduate exercises and more advanced mathematics, including the prime number theorem, using a map of statements and their dependencies.

That work translated established mathematics into a form software could check. It also produced definitions and intermediate results that other researchers could reuse. The example shows why the value of a workshop can extend beyond a single paper. A shared library reduces the amount of foundational work later projects have to repeat.

The foundation described that earlier work in an article dated July 22, 2026. It offers a practical model for turning an intensive research meeting into materials that remain useful afterward.

OpenAI’s October 6 release provides a more recent reason to invest in review. It published manuscripts, Lean formalizations and information about the research process from an internal model. ByteForward’s earlier coverage explains how to examine that collection.

The repository’s October 7 revision record documents three withdrawn manuscripts after a sign error affected a construction and dependent work. It also records repairs and clarification in other papers. The lesson for a funded review is to track the exact version being examined and its dependencies.

What a checked proof can establish

Lean’s technical documentation describes a small trusted kernel that checks formal proof terms against its logical rules. That gives a precise target for verification. Researchers still have to establish that the formal statement captures the mathematics they intended to claim.

The official proof validation guide separates a valid proof from the meaning of its theorem statement. A reassuring editor indication can coexist with incomplete arguments in dependencies. Reviewers therefore need to examine axiom dependencies, imported definitions and the scope of what was checked.

For stronger checks, Comparator can compare a proof against a theorem statement prepared in a trusted environment and use external checkers. That still depends on a correct challenge statement and sound checking infrastructure. A successful run does not determine whether the result is new, important or explained well enough for humans to use.

Scope can be narrower than a paper title suggests. OpenAI’s Comparator instructions explicitly identify some configurations as checks of supporting results rather than the corresponding main theorems. Reporting that distinction is part of reproducibility. A review should say exactly which statement was tested, instead of describing an entire paper as verified.

ByteForward has examined the published descriptions and records for this article. We have not built these proof artifacts or independently verified the mathematics.

Reproducing a proof and reproducing a discovery

A reproducible check needs a usable software environment. Lean’s build documentation identifies the toolchain used to check source files and Lake’s role in tracking dependencies. In practice, reviewers should preserve the repository revision, toolchain version and dependency state alongside the commands and results. That gives another group a defined starting point.

Recreating the original discovery is a different task. OpenAI’s repository description says the work used an unreleased internal model and includes results at different stages of verification. Published proof artifacts can enable inspection without providing the same model or reproducing the search that generated them.

The advisory group also calls for broad, equitable model access and objects to advanced mathematical research on proprietary systems unavailable to the wider community. Its recommendations ask labs to disclose the model, prompts, computation and process behind released results. The new partnership should not be read as an endorsement by that group.

The next useful evidence will be the application terms and the work participants make available. A review that explains an assumption, repairs an argument or makes a proof easier to reuse can justify its cost. Those are practical outcomes worth measuring.

Illustrative photograph of scientific study, published in 2025. Photo by Vitaly Gariev on Unsplash under the Unsplash License.