A Millennium Prize problem fell this month — not to a mathematician at a chalkboard, but to a swarm of AI agents racing a rival team by half a day.
On September 8, OpenAI announced that a deployment of roughly 10,000 autonomous AI agents, working for 88 hours and exchanging nearly 5 million messages, had produced a Lean-verified proof that solutions to the Navier–Stokes equations — the equations governing fluid flow — can "blow up," reaching infinite speed in finite time. It's one of the Clay Mathematics Institute's seven Millennium Prize Problems, each carrying a $1 million reward. Twelve hours earlier, a competing effort by NYU's Tristan Buckmaster and Harvard's Levent Alpöge, working with Anthropic's Claude, had posted proofs of the related zero-viscosity Euler case, apparently accelerating their release after hearing OpenAI was closing in. Both efforts lean on a mathematical strategy Diego Córdoba and Luis Martínez-Zoroa worked out years earlier by hand — Princeton's Charles Fefferman called them "the heroes of the story" — and the AI-generated write-ups weren't uniformly clean: Buckmaster described one early draft as "the most horrendous I have ever read." Mathematicians are still sorting out how much credit belongs to the humans whose decades of work the agents were built on.
Why it's hereThe math is real and dramatic on its own, but the more interesting story is procedural — a scramble between two AI-assisted teams, with credit, cleanliness, and attribution all still unsettled.
The world's most sensitive dark-matter detector saw one unexplained event. The team's response was to say, clearly, that it isn't enough.
LUX-ZEPLIN (LZ), a 10-tonne tank of liquid xenon buried nearly a mile underground at the Sanford Underground Research Facility in South Dakota, spent 220 days between March 2023 and April 2024 watching for the faint recoil a WIMP — a leading dark-matter candidate — would leave behind. Combing through that data, the roughly 250-scientist collaboration found exactly one event sitting in the energy range and topology where such a recoil would be expected, corresponding to a WIMP mass of at least 200 GeV — more than 200 times a proton's mass. It registers at 2.6 sigma, meaning roughly a 0.5% chance it's a fluke of the background, well short of the 5-sigma bar physics requires before calling something a discovery. "We're very intrigued to see this event," said LZ spokesperson Rick Gaitskell, "but we are not claiming to have seen dark matter" — and Brown physicist JiJi Fan noted the signal doesn't even fit the simplest WIMP-interaction model cleanly, leaving inelastic scattering and other explanations on the table. LZ is still running, so more data should either strengthen or dissolve the signal within a year or two.
Why it's hereThe signal itself may well turn out to be nothing — the value here is watching a large collaboration handle a single tantalizing data point honestly, with the statistics stated plainly rather than teased.
The industrial reaction that turns natural gas into syngas has run on nickel for decades. It turns out the nickel wasn't doing what chemists thought.
Partial oxidation of methane — the reaction that converts natural gas into syngas, the CO/H2 feedstock behind fertilizer, methanol, and synthetic fuel production — has long been assumed to run on metallic nickel nanoparticles. A team at the Dalian Institute of Chemical Physics, led by Tao Zhang, used in-situ spectroscopy and modeling to show that's not quite right: under reaction conditions, nickel oxide surfaces reorganize into a small nickel-oxygen cluster (a "Ni-O-Ni₄" motif) that turns out to be the actual active site, cutting the energy barrier for breaking methane's C–H bond to 12.5 kcal/mol, well below the 38.5 kcal/mol on intact nickel oxide or 15.7 kcal/mol on metallic nickel. Built around that insight, their catalyst needed only 0.8% nickel by weight — roughly a tenth of conventional loadings — yet still hit 92% methane conversion and 87% syngas selectivity at 650°C. Because syngas production runs at enormous industrial scale and nickel is a volatile-priced strategic metal, a mechanistic correction like this one has real cost implications, not just an academic footnote.
Nature Catalysis · Dalian Institute of Chemical Physics · Aug 14, 2026 · nature.com
§
Why it's hereA rare case of a textbook mechanism getting corrected on an industrially important, massive-scale reaction — the kind of result that changes catalyst design, not just a lab curiosity.
The same genetic adaptations that let bats shrug off viruses may also be why they barely age or get cancer.
Bats in the genus Myotis show an enormous, size-independent spread in lifespan — one banded Brandt's myotis was recaptured 50 years after tagging, while a related species lives only about seven — despite bats carrying unusually high viral loads as reservoirs for coronaviruses and more. A team led by Juan Vazquez and Peter Sudmant at UC Berkeley sequenced and compared genomes across eight Myotis species and found that longevity tracks with positive selection in cancer-suppression and DNA-repair pathways, alongside a distinctive damage response: rather than trying to repair severely damaged DNA — itself a cancer risk — cells in the long-lived Myotis lucifugus decisively kill themselves off instead. "The longest-lived bat in North America decides 'I can't save this ship' and immediately switches gears to prioritize killing off the cells that are damaged," Vazquez said. The genomes also showed bats using different immune strategies against DNA viruses versus RNA viruses, including a segregating duplication in the antiviral gene PKR — evidence, the authors argue, that antiviral and anti-aging adaptations evolved together rather than as separate evolutionary problems.
Why it's hereNot just another "bats are weird" story — a concrete cellular mechanism, kill rather than repair, that ties cancer resistance and viral tolerance to the same evolutionary root.
A single CRISPR edit to one liver gene lowered cholesterol and triglycerides — and, unusually, the effect hadn't faded twelve months on.
CTX310, an in-vivo CRISPR-Cas9 therapy from CRISPR Therapeutics, permanently disables ANGPTL3, a liver gene that normally restrains how fast the body clears fat from the blood — mimicking a rare natural mutation long known to protect against heart disease. In a Phase 1 dose-escalation trial of 15 adults with hard-to-treat lipid disorders, the highest dose cut LDL cholesterol by a mean of 52.5% and triglycerides by a mean of 47.8% at 12 months, with no sign of the effect fading — durability that stands out, since most lipid drugs have to be taken indefinitely. Three patients had Grade 2 infusion reactions and one had an allergic reaction, all resolved, with no serious treatment-related adverse events reported. "It is encouraging that there were no serious safety events related to CTX310 in the trial and in the year following treatment," said Cleveland Clinic's Dr. Luke Laffin, the paper's first author. The caveats are real: this is 15 patients, uncontrolled, from a narrow geographic pool, and because the edit is permanent, the FDA is requiring 15 years of long-term follow-up before anyone can be confident it's safe over a lifetime.
Why it's hereA genuine test of "one-shot gene editing" for a common chronic condition rather than a rare disease — promising durability, paired honestly with just how far off real confidence still is.
A researcher argues journals should stop pretending peer reviewers aren't already using AI, and start building rules for how they should.
Most journals currently ban reviewers from using AI tools to help evaluate manuscripts, largely over two concerns: exposing confidential, unpublished work to public AI systems, and diluting a reviewer's own accountability for their judgment. In a Nature correspondence, Tsinghua University's Xuegong Zhang argues the ban is the wrong response to a real problem — submission volumes are rising, manuscripts are getting more complex, and journals are struggling to find reviewers who can turn work around in time. His point isn't that AI assistance is risk-free; it's that a blanket prohibition "address[es] concerns imperfectly and [is] hard to monitor," since reviewers can and likely do use AI quietly regardless of the rule. He argues for frameworks that build in real safeguards — confidentiality-safe tools, disclosure requirements — instead of a policy that mostly just pushes the practice underground.
Nature (Correspondence) · Aug 25, 2026 · nature.com
§
Why it's hereA stakeable, unresolved policy question — ban vs. regulate — that every academic reader publishing or reviewing right now has a personal position on.