Machine Learning Takes On the Hardest Math on Earth — and the Results Are Stunning

Machine learning just had one of the most extraordinary 24 hours in its history. On November 20, researchers posted a claimed proof of a million-dollar Millennium Prize problem — built with heavy help from AI models. The same day, another mathematician said an AI helped him prove the Riemann hypothesis. And a respected evaluation lab warned that AI capabilities are now improving faster than our ability to measure them. Machine learning is no longer just writing code and answering questions. It is reaching into the hardest problems in mathematics — and the field is struggling to keep up with itself.

The Navier-Stokes Claim: A Million-Dollar Problem, AI-Assisted

The headline story comes from NYU mathematician Tristan Buckmaster and Anthropic researcher Levent Alpöge, who posted a 79-page paper on November 20, 2026 claiming a finite-time blowup proof for the three-dimensional Navier-Stokes equations with no external force, as reported by Emergent.sh. The Navier-Stokes equations describe fluid flow and have resisted complete mathematical understanding for over a century. The specific claim is that solutions can blow up — meaning certain physical quantities become infinite in finite time — and it targets a problem worth $1 million from the Clay Mathematics Institute.

What makes this paper a machine learning story is how it was made. The proof includes a 600,000-line Lean 4 formalization created with heavy assistance from Anthropic and OpenAI models, including Codex. The researchers openly acknowledge their reliance on large language models to generate both the paper text and the formal verification code. The Lean 4 formalization spans approximately 2,270 files built on the Mathlib library, representing months of machine-generated formal verification work. Buckmaster himself described the write-up as “unpolished and heavily LLM-assisted” in a Mastodon post hours after publication.

There is a catch, and it matters. The critical numerical certificate the proof depends on was not released — only a cryptographic hash promising future availability. Without that certificate, independent mathematicians cannot verify the central computational claim. Scientific American’s coverage noted that no outside mathematician had publicly vouched for the result at the time of publication. The hash is a commitment mechanism, but it does not substitute for actual verification.

The timing reflects how AI tools are compressing research cycles. What might have taken years of manual proof-writing and formalization was accomplished in months — though at the cost of polish and immediate verifiability. The authors’ candor about their methods, including listing which AI models contributed to which sections, sets a transparency standard that may become necessary as LLM-assisted research grows more common.

A Second Claim the Same Day: The Riemann Hypothesis

The Navier-Stokes paper was not even the only headline. As AI.info’s daily briefing for Saturday, November 21, 2026 reports, Anthropic mathematician Levent Alpöge says he and Claude found a proof of the Riemann hypothesis and formalized their argument. The proposed proof has not been independently verified. Alpöge said the arguments were formalized without assumptions, while The Verge says the proposed solution is presented in Lean.

Two claimed proofs of century-old problems in a single day, both built with machine learning systems. Even if neither claim survives peer review, the pattern is unmistakable: ML models have moved from assisting with routine proofs to participating in work at the very frontier of mathematics.

METR: Capabilities Are Outrunning Our Measurements

While the math world digests those claims, the evaluation lab METR published a warning that should concern anyone who builds on machine learning. According to a Dosa AI Tools briefing on November 21, 2026, METR published capability evaluations of public models up to Opus 5.5 and GPT-6 Astra, saying it remains significantly uncertain what capabilities are needed for drastic AI research and development acceleration — and how fast they improve. Current public evidence cannot rule out such acceleration in the next couple of years.

AI.info’s daily briefing adds the detail that drives the point home: METR says tested models usually matched or exceeded skilled humans on coding and optimization tasks, and estimates time horizons are doubling roughly every 3.6 months. In other words, the length of task a model can handle autonomously keeps doubling every few months — and the lab whose job it is to measure this says the measurements themselves are getting harder to make. Teams using benchmark scores to pick models should treat them as weaker signals than a year ago.

Models That Improve Themselves — and Fit on a Laptop

Two more developments from the same day show how fast the ground is shifting. Researchers report that an open 27-billion-parameter language model edited its own weights and improved by 6.6 percentage points on unseen problems, according to AI.info. Self-improvement is no longer a thought experiment; it is a preprint result.

At the same time, the tools are getting smaller and more accessible. Thinking Machines Lab released Inkling 2 Mini on November 20, an open-weight model with an approximately 20GB memory footprint — small enough to run on a laptop — along with a research variant to study how repeated passes through model layers affect performance and oversight. And Baseten’s Base Labs released bison-1, a 31-billion-parameter dense model, and groundhog-1, a 26-billion-parameter mixture-of-experts model with about 4 billion active parameters, both built on Gemma 4 with a 128k context window. Bison-1 ranks first of 19 models on the Artificial Analysis Decision Index v2.1 at 76, and both are available on Hugging Face. NVIDIA also opened Nemotron for Science, combining scientific applications of its Nemotron 3.5 Super VL open model, post-training recipes, and LitmusBench v0.2, an open dataset of nearly 50,000 chemistry reasoning questions.

The picture is a field expanding in every direction at once: frontier models tackling Millennium Prize problems, open models small enough for a laptop, and scientific models with public benchmarks for chemistry research.

What This Week Means for Machine Learning

The week’s lesson is about verification. Machine learning can now produce a 600,000-line formal proof, a claimed Riemann hypothesis solution, and a self-improving model — but the Navier-Stokes proof cannot be checked until its numerical certificate appears, and the Riemann claim awaits independent verification. The same gap showed up in the evaluation world: METR’s warning is, at its core, a verification problem too. The models are moving faster than the tests designed to measure them.

For teams building with machine learning, the practical takeaway is to build verification into the workflow from the start. Demand the artifacts, not just the claims. Check the certificate, run the benchmark yourself, and test on your own data. The models are powerful enough now to produce work that looks complete — the discipline of checking it has never mattered more.

Machine learning spent years proving it could help with the easy problems. This week it took on the hardest ones. Whether the proofs hold up or not, the field will never look the same.

The Bottom Line

On November 20, AI-assisted researchers claimed proofs for the Navier-Stokes blowup problem and the Riemann hypothesis, with a 600,000-line Lean 4 formalization behind the first — but the key numerical certificate is still withheld, so neither claim is verified yet. METR warns that model capabilities are doubling in time horizon every 3.6 months, faster than our measurements can track. And open models keep getting stronger and smaller, from a self-improving 27B model to a laptop-sized Inkling 2 Mini. Machine learning has reached the frontier of mathematics — now the field has to prove it can verify what it builds.

Sources: Emergent.sh, “AI Models Help Prove Navier-Stokes Blows Up in Finite Time,” November 21, 2026; AI.info, “The Daily · Saturday 21 November 2026 — the day in AI”; Dosa AI Tools, “METR warns AI R&D gains are outpacing measurement,” November 21, 2026; KimKJ.com AI Daily Briefing, November 21, 2026.

Edited by Muse @QUE.COM
Website: https://QUE.COM Intelligence
Sponsored by: https://MAJ.COM AI Autonomous
Never miss a customer call again — MAJ.COM’s AI receptionist answers every call, texts customers back instantly, and books the appointment for you. See how it works at https://MAJ.COM


Discover more from QUE.com

Subscribe to get the latest posts sent to your email.

Leave a Reply

Connect with

Discover more from QUE.com

Subscribe now to keep reading and get access to the full archive.

Continue reading