A weekly notice from the margins of math and science
Marginalia
the good stuff, allegedly
Issue №8·20 October 2026
AI shows up twice this issue — in different fields, doing different jobs. One is a chatbot company's model putting a real, still-being-checked dent in one of math's most famous open problems. The other is a much more ordinary AI tool quietly doing the grunt work of chemistry: sorting through thousands of reaction steps to find one humans had missed. Neither is "AI did the whole thing," and we've tried to be honest below about exactly where the line is.
A 167-year-old guess about how prime numbers are scattered carries a $1 million prize and has beaten every mathematician who's tried it. This month, an AI may have proven a meaningful chunk of it.
In 1859, mathematician Bernhard Riemann guessed that a mathematical object called the zeta function has all of its interesting "zeros" — points where it equals exactly zero — lined up along a single line. If true, that would pin down exactly how evenly prime numbers are spread out among all the whole numbers, which is why it's one of math's six unsolved Clay Millennium Prize problems, worth $1 million to whoever proves it; mathematicians have checked the first 12 trillion zeros by computer and found the pattern holds every time, but a computer check isn't a proof. On October 6, 2026, OpenAI released a batch of AI-generated papers, including one claiming to prove a weaker version called the quasi-Riemann Hypothesis: that those zeros are confined to a narrower band than previously known, even if not yet pinned to Riemann's exact line. Number theorists including Maksym Radziwiłł at NYU and Alexander Dunn at Georgia Tech have spent the weeks since checking the argument line by line, and Stanford's Jared Duker Lichtman called it "the most significant AI-for-math achievement so far, by quite a margin" — though Princeton's Peter Sarnak says he won't call it a real theorem until humans have verified every detail by hand, and the full, original Riemann Hypothesis remains as unsolved as it was in 1859.
Quanta Magazine, reporting on papers released by OpenAI · Oct 19, 2026 · quantamagazine.org
§
Why it's hereAI tools have been nibbling at the edges of serious math for a couple of years, but this is the first time one has plausibly — not certainly — landed a real blow on one of the field's most famous open problems, and working mathematicians are still arguing in public about whether to believe it.
Werner Heisenberg thought his wartime reactor just needed a little more uranium to go critical. Two surviving cubes from that machine, examined 80 years later, show he was off by a factor of two.
During World War II, the United States raced to build the atomic bomb partly out of fear that Nazi Germany was doing the same. Germany's program, led in its final stages by physicist Werner Heisenberg, built experimental reactors out of natural uranium metal cubes suspended in heavy water, which slows down neutrons enough for a chain reaction to sustain itself. The last of these, called B8 and built in 1945, was dismantled by American forces before it ever reached "criticality" — the point where a reaction keeps itself going without outside help — and Heisenberg always maintained B8 had been close, needing only a modest boost in uranium. A new study led by Timothy Koeth, combining physical analysis of two of B8's actual surviving uranium cubes with wartime technical records and correspondence, found Germany would have needed roughly twice the uranium it had and more than twice the heavy water, and that a graphite-based backup design wouldn't have worked either, because boron contamination in the only graphite Germany could get its hands on soaked up too many of the neutrons it needed. Published in the journal *PNAS Nexus*, the analysis is some of the clearest physical evidence yet that the German program was never a real threat, more than eight decades after the war ended.
ScienceDaily, reporting Timothy Koeth et al., PNAS Nexus · Oct 16, 2026 · sciencedaily.com
§
Why it's hereIt's rare to get hard, physical evidence settling an 80-year-old historical argument, and this one works as both real nuclear physics — what "critical" actually requires — and as a genuinely dramatic piece of WWII history.
Turning carbon dioxide into usable fuel on a copper catalyst seemed simple enough to model by hand. A machine-learning sweep of thousands of possible reaction steps found a whole pathway nobody had included.
One way to recycle carbon dioxide instead of just venting it into the air is to react it with hydrogen on a copper catalyst to make methanol, a liquid fuel and industrial chemical — but predicting which of the many possible reaction steps actually happens, and in what order, is hard enough that chemists have long relied on small, simplified models of the process. Researchers at the Indian Institute of Science, led by Anand Mohan Verma and Ananth Govind Rajan, built a far bigger model instead: they ran detailed quantum calculations on 152 individual reactions, trained machine-learning tools to estimate the energy cost of thousands more, and let automated software search through every plausible one-step transformation among 105 chemical fragments sitting on the catalyst's surface. That search expanded the known reaction network from 152 steps to 9,389, and revealed that hydrogen often transfers as a whole molecule (H2) rather than as two separate atoms, a pathway the smaller, hand-built models had missed entirely. The expanded model predicted roughly 40 times more CO2 conversion than the old one and correctly picked methanol, rather than formic acid, as the main product, matching outside lab tests run in India and Singapore. It's still a computer model, not a working factory process, but it gives chemists a genuine, specific lead on which copper catalysts to try building next.
ScienceDaily, reporting Anand Mohan Verma, Ananth Govind Rajan et al., Nature Communications · Oct 13, 2026 · sciencedaily.com
§
Why it's hereClimate-fix chemistry headlines get hyped a lot, and the honest version of this one is still interesting on its own: a smarter simulation found a mechanism nobody had thought to check, which is exactly how better catalysts actually get designed — a computational lead, not a finished fuel plant.
Severely bleached coral fragments that touched a healthy, genetically identical piece for just a few days fully recovered within seven weeks. The ones kept a few millimeters apart died.
When ocean water gets too warm, corals expel the colorful algae living inside their tissues that supply most of their energy — a process called bleaching that, if it lasts too long, kills the coral outright. Stanford researchers led by oceans PhD student Danny Collins noticed, almost by accident, that a bleached coral fragment touching a healthy piece from the very same original colony started regaining its color, so they tested it on purpose: ten matched pairs of fragments, one bleached and one healthy from the same coral animal, with five pairs pushed together and five kept a few millimeters apart. Every touching pair fully recovered within seven weeks, with algae visibly crossing over through newly formed tissue channels between the two fragments within just three days; every separated pair died. Once enough algae had moved across, they multiplied inside the recovering coral roughly ten times faster than researchers have seen algae grow on a wild reef, and once that initial transfer happened, the two fragments didn't even need to stay in contact. The technique only works between genetically identical fragments and hasn't been tried outside the lab, but it adds a genuinely new option, alongside breeding heat-tolerant corals and raising them in nurseries, for helping reefs survive the weeks after a marine heat wave.
ScienceDaily, reporting Danny Collins et al., Stanford University / Science Advances · Oct 19, 2026 · sciencedaily.com
§
Why it's hereCoral bleaching stories are almost always just bad news. This one has an actual controlled experiment, a concrete number (seven weeks, every single touching pair recovered), and a hopeful — if still lab-only — fix.
Gary Keblish's melanoma was removed six years ago. He's still teaching, partly because he was one of the first people to get an mRNA vaccine built from the specific mutations in his own tumor.
Moderna and Merck's intismeran autogene is a personalized cancer vaccine: an algorithm scans a patient's own tumor, picks out up to 34 of its specific mutant proteins, and builds an mRNA shot that trains the immune system to recognize and attack cells carrying those exact mutations, given alongside Merck's immunotherapy drug Keytruda. In a five-year follow-up of an earlier mid-stage melanoma trial, presented at the 2026 ASCO meeting, patients who got the combination had a 49% lower risk of their cancer recurring or killing them, and a 59% lower risk of it spreading to distant organs, compared to Keytruda alone. A larger Phase 3 trial, called INTerpath-001, confirmed the benefit still held at a planned interim check announced in August 2026, hitting both of its main goals in patients with high-risk melanoma that had just been surgically removed. Gary Keblish, a public school teacher diagnosed with melanoma in 2019, was randomly selected into the vaccine arm of that earlier trial; he says the nerve-damage side effects were the hardest part, but he's glad he got it, and his tumor has been gone for six years while he's back in the classroom. Moderna and Merck are now preparing to file for regulatory approval and are already testing the same personalized approach against lung, kidney, bladder, pancreatic, and gastric cancers.
STAT News, reporting the 2026 ASCO Annual Meeting and INTerpath-001 trial data · Oct 15, 2026 · statnews.com
§
Why it's here"A vaccine built from your own tumor's mutations" sounds like science fiction, but the risk-reduction numbers here are real trial data, not just a hopeful anecdote — and the patient underneath those numbers is a solid, concrete reason to care.
Sources this issue
Quanta Magazine, reporting on AI-generated papers released by OpenAI — math
ScienceDaily, reporting Timothy Koeth et al., PNAS Nexus — physics
ScienceDaily, reporting Anand Mohan Verma, Ananth Govind Rajan et al., Nature Communications — chemistry
ScienceDaily, reporting Danny Collins et al., Stanford University / Science Advances — biology
STAT News, reporting the 2026 ASCO Annual Meeting and INTerpath-001 trial data — medicine