Forecasting Research Institute × LEAP

Probability AI Solves Three Millennium Prize Problems

Probability AI Solves Three Millennium Prize Problems

Experts put a 45% chance on AI helping solve at least three of the five remaining problems by 2030, and 80% by 2040.

Hover to highlight a group, click to show or hide it

Our Analysis

How do we interpret these results

All groups consider it very unlikely that AI solves, or substantially assists in solving, at least three of the five remaining problems by the end of 2026, with medians between 1.5% and 3%, but their forecasts rise substantially by 2030 and further by 2040. For 2030, superforecasters and AI risk experts give medians of 50% and experts 45%, which are statistically indistinguishable, while the public is significantly lower than all other groups at 20%. For 2040, superforecasters again give the highest median forecast at 88%. Experts and AI risk experts both give 80%, statistically indistinguishable from superforecasters, while the public is again significantly lower than every other group at 49%.

Forecasts by group

Median forecasts from each group

group202620302040
Experts
3%
45%
80%
Superforecasters
1.5%
50%
88%
Public
3%
20%
49%
AI Risk Experts
2%
50%
80%

Click any cell to compare

Rationale analysis

The main lines of reasoning driving the spread in estimates

AI's trajectory in math

Most respondents predicting fast progress point to the pace of general AI progress in math, the rise of AI agents working in parallel, the September Navier–Stokes proof and OpenAI's release of other solved frontier-math problems, or all of the above, as a sign that Millennium-level mathematics is within AI's reach: “I don't think these math problems need fully new ways of thinking. They can be solved by brute force using a tree of agents.”; “The Navier-Stokes result says that solving Millennium prizes is a good deal easier than I used to think.” Others acknowledge that general progress has been fast but tend to downplay the Navier-Stokes breakthrough: “OpenAI's result is impressive, but it exploits a loophole in the official problem statement…and it builds on ground humans had already prepared extensively.”; “Creativity is not the same as brainpower/compute power.”

Difficulty of the remaining five

Many skeptics argue that the remaining problems are quite different from and likely more difficult than Navier–Stokes: “These five problems are not interchangeable [in terms] of difficulty, and success on one does not necessarily imply that the others become straightforward.”; “The remaining five are mostly 'prove a universal statement' problems with no known viable program (P vs NP faces relativization/natural-proofs/algebrization barriers; RH has no credible attack; Hodge/BSD need new theory).” One adds, “There's no guarantee that all of them can be solved.” Fast-progress respondents tend to single out problems that are rumored to be within reach: “Hodge and BSD [Birch–Swinnerton-Dyer] seem like gimmes (both saw significant progress in the OpenAI drop).”

The 2026 window and verification

Slow-progress respondents often point out that a proof would need to be found, released, and vetted to count toward a 2026 resolution. “Three problems is an absurd number for 2.5 months,” writes one. Another calls confirmation by year-end “nigh impossible.” Fast-progress respondents, however, tend to see the expert-panel route as viable: “The panel criterion is the only realistic path in 2026, and FRI's ruling that the Navier-Stokes claim meets it means a public, credible, ideally Lean-verified proof is enough.” Some suspect results are already waiting: “I think there's a decent chance one or more companies is sitting on a solution already.”

Three of five

Most respondents agree part b is a much higher bar, but differ over the degree to which success in solving one of the remaining five problems will presage success in solving the others. Low-progress respondents tend to stress the uniqueness of the individual problems: “The mathematical machinery needed to approach the Hodge Conjecture may have relatively little to do with resolving P versus NP.” Whereas fast-progress respondents typically expect jumps in general capability to result in a domino effect: “If it solves another, it will soon solve them all.”

Bragging rights versus the backlash

Fast-progress respondents often stress that labs have clear reasons to keep trying: “There's a literal prize for solving them, but I think the prestige, publicity, and ability to demonstrate their models' capabilities will be an even greater incentive.”; “Perfect as a showcase of AI prowess before an IPO.” Others emphasize that cost and mathematicians' hostile reaction could slow things down. “Given that Navier-Stokes already fell, the news value or propaganda value of another one is unclear. Might not be worth all the compute?” writes one. Another adds, “One factor against is the very negative reception from the mathematical community to Navier-Stokes.” Others speculate mathematicians might share less, “given the rumors of the AI Labs allegedly stealing ideas and methods from their users.”

“Substantially assist”

Many fast-progress respondents argue that any future solution will involve AI, so the question comes down to whether the problems get solved at all: “I have a hard time imagining at this point that any solution will not substantially involve AI.” A few slow progress respondents question this assumption, with one writing, “Experts must judge that an original, important part of the solution likely could not have been produced without AI. Widespread use of AI in mathematical research would [still] leave this question open.”

Rationale examples

High-quality rationales from forecasters across the range

Fast-progress respondents

I have given a lot of thought to exactly this question over the past two years, and even more so since Navier Stokes was resolved. I predicted IMO gold in 2025 based on 2024 performance and the trendlines in the mathematical capabilities of frontier models. Also since 2024, I have been predicting that all remaining Millenium Problems will be solved by AI before 2030. This is actually a somewhat conservative prediction that allows for the historical exponential trajectory to slow down in the near future. Today, existing frontier AI systems are already strictly superhuman at mathematics in almost every way, and the Millennium Problems no longer look so formidable. OpenAI's internal 'Bel' model may be sufficiently capable to fully resolve Hodge and BSD later this year. Additional post-training on top of the hundreds of recently solved problems could result in a future 2026 model capable of fully solving Riemann as well. Yang-Mills is stubborn and P vs. NP is treated as almost mystical vis a vis its preeminent status in complexity theory; those two are the likeliest holdouts.

What's particularly interesting to me is the emergence of coordinated agent systems capable of exploring multiple approaches to difficult problems, investigating intermediate conjectures, consulting existing literature, and iteratively refining proposed solutions. This is a fundamentally different computational arrangement from asking one model to contemplate the Riemann Hypothesis for several minutes and hoping that enough tokens will eventually produce a Fields Medal. We're beginning to build systems that can sustain mathematical investigations over extended periods, with multiple independent lines of inquiry and tools capable of checking at least some of the resulting work, and I suspect the implications of that change are not yet fully appreciated.

Some rough sources of intuition: 1. Human base rate on the original 7 was 1 problem in ~26 years; 2. There are currently around 90k human mathematicians, and GPT-6 gives a BOTEC of 2.5M researcher-years spent on math research between 2000-2026; 3. Median US research mathematician salary is ~$130k USD/year. Accounting for skew and selection where more expensive researchers are working on harder problems, let's call the average $200k; 4. If 1/500 researcher-years were spent working towards Navier Stokes (including subproblems), that would be ~$1B spent on human salaries; 5. OpenAI then solved it for $15M. So their AI had a cost efficiency perhaps 67x better than human researchers; 6. Cost for an equivalent level of AI performance falls ~13x/year, so running models as capable as today's frontier will be (at least) several OOMs cheaper by 2030; 7. Even if relative cost efficiency stays fixed (it's surely increasing), that might suggest that you could match all human mathematical effort from 2000 to 2026 with a few hundred thousand dollars by 2030. The net effect is that, very likely, there will be more effort from AI than all mathematicians in history by the end of the decade.

...seeing the recent results made all the mathematicians in my [circle] freak out and everyone is scrambling to get their agent swarm and compute credits. So if these problems are solved AI will be involved. I expect AIs to become smarter than humans in all domains by 2040.

Slow-progress respondents

The Navier-Stokes result is less of a template than it looks. It was a blowup construction sitting on years of human groundwork (Cordoba / Martinez-Zoroa, Buckmaster et al. on Euler), where the remaining gap was a huge but searchable, verifiable construction that 10k agents x 88h plus Lean could close.

The Navier-Stokes result shows AI can now finish off a hard problem, but that one had a human research program already close to the answer, and the credit is still being argued over. The other five (Riemann, P vs NP, Hodge, BSD, Yang-Mills) don't have anything like that close to done, and some need whole new ideas. 2026 is almost over and a claim would also need expert sign-off, so that's very unlikely.

It will be more difficult than we think now. Many of the problems will require 'extrapolation', while AI has so far been good at 'interpolation' type of maths problems - combining theories from different topics & deploying massive parallelism to scale search, rather than coming up with…completely new language systems, for example.

Three of five is a much stronger claim and I treat it as close to requiring a general mathematical reasoner rather than a sequence of lucky fits. [Multiple solutions implies] that the framework-invention problem gets solved…not the probability that compute keeps scaling; I do not think these are the same question, and the field's history is that the hard conjectures fall to new ideas, decades apart, not to more effort.