Dear Readers, 

Most of us are getting used to the feeling of terror and awe as AI does yet another thing better than we ever could. But mathematicians are experiencing this on another level entirely, as AI models topple one mathematics challenge after another. What to do? Last month, over 3,000 mathematicians signed a declaration intended to set guidelines for how AI and the math profession should co-exist going forward. And they have a warning for other academic disciplines: You could be next. 

Also in the issue: 

Thanks for reading, 

Danielle Mattoon 
Executive Director

Subscribe

Subscribe to our newsletter and be kept up to date on upcoming Aventine projects

While many of us have watched with a mixture of panic and awe as AI gobbles up fundamental parts of our jobs, mathematicians are seeing the very rationale for their existence threatened as machines topple one mathematical challenge after another. 

Progress has been swift. This time last year, artificial intelligence models joined the ranks of International High School Math Olympiads. In the months since, they have solved math problems that had stood unsolved for decades, using reasoning that their human peers consider worthy of publication. 

For some who have devoted their lives to advancing the frontier of mathematical understanding by developing proofs and theorems, it’s a world-rattling experience. “The field is going to change, and these tools are going to change the field, whether people are on board with it or not,” said Bryna Kra, a math professor at Northwestern University and former president of the American Mathematical Society. 

What to do? The question is the basis of the recently published Leiden Declaration on Artificial Intelligence and Mathematics. Signed by over 3,000 people, including math heavyweights such as Terence Tao from UCLA and Peter Scholze, director of the Max Planck Institute for Mathematics, the document is a stark warning about how AI could undermine the field of mathematics, and a 23-point plan aimed at mathematicians, math institutions and policymakers for preserving math as a human-first endeavor. 

In presenting mathematicians as a united front, the document belies a profession struggling with disruption in real time. Behind the scenes, some mathematicians are embracing new AI tools, eager to advance mathematics in any way possible while others are uneasy about handing tasks to systems that can’t be fully understood. Questions about how the field will adapt — or be forced to adapt — to accommodate AI, how human accomplishment can be preserved, where a line should be drawn in trusting what an AI produces and how to work productively with the AI labs are all up for grabs. And while the field of math is in some ways uniquely exposed to AI, some mathematicians believe it is a bellwether for the impact AI will have on other disciplines and that their response will help shape the way other sciences adapt to the technology. 

AI vs. mathematicians

Achievements of AI in mathematics have been gradually stacking up for months. In the summer of 2025, models from Google DeepMind and OpenAI achieved gold medal status at the International Mathematical Olympiad, a contest for the world's most mathematically gifted high school students. Then last winter, models started biting off low-hanging research problems, notably some of those formulated by Paul Erdős, a prolific Hungarian mathematician who died in 1996, leaving behind more than a thousand deceptively simple yet unsolved questions, many of which remain so.

In recent months, AI models began solving problems many mathematicians say constitute PhD-level research. Google DeepMind’s Aletheia model solved a problem in arithmetic geometry that had eluded human researchers. More recently, a general-purpose AI system built by OpenAI disproved a conjecture (a mathematical theory that hasn’t been proven right or wrong) with reasoning that some mathematicians considered worthy of publication in a major journal. The way the OpenAI model approached the problem, combining ideas from disparate mathematical disciplines, was seen by some as a paradigm shift, illustrating how language models can become experts in “everything all at once," said Kevin Buzzard, a math professor at Imperial College London. 

"You can basically solve conjectures within a day that were lying low for months or years," said Bartosz Naskręcki, a math professor at Adam Mickiewicz University in Poland and a co-author of the Leiden Declaration who also collaborates with OpenAI. He described the overall rate of progress as "ridiculous." 

But AI is not good at everything when it comes to math, and recent results don’t impress all mathematicians. Ravi Vakil, a math professor at Stanford University and the current president of the American Mathematical Society who has worked with Google DeepMind, said that "[AI] can prove interesting things, but every single time it's being pointed in that direction by people." Aimed at long, complicated problems, he added, it quickly starts to get things wrong. Naskręcki, meanwhile, pointed out that the recent OpenAI result was an example of "cherry picking," a single success from hundreds of attempts. He also downplayed the complexity of the work, saying that “it wasn't some kind of rocket science.”

There are fault lines running through the community about what happens next. One is whether, in the future, mathematicians need to be able to understand the work that AI creates. Proponents of AI argue that we should embrace the expansion of the mathematical frontier regardless of whether a human comprehends it; critics contend that if a human cannot understand an AI-generated proof, we may never truly know if it is correct. "If you don't know why it's true, you don't know what the next interesting question is," said Vakil. Buzzard imagines a future in which an AI produces a giant proof of the long-unproved Riemann Hypothesis and an automated system verifies it — at which point "some people think ‘This is great,’ some people think ‘This is a disaster.’" It is possible that the field will bifurcate into camps that are comfortable with AI making advances they don't understand, and those that aren't. 

The Leiden Declaration tries to thread the needle, articulating the concerns about AI that mathematicians largely share, arguing that researchers should be transparent about their use of AI, take responsibility for correctness and choose their tools carefully. Other recommendations are, for now, more aspirational: that policymakers regulate the AI industry, and invest in public computational infrastructure.

What do AI labs want from math? 

The relationship between mathematicians and AI labs has been one of the more difficult areas on which to reach consensus. Across many conversations for this story, there was unease about the power the AI companies wield, alongside a desire to experiment with what they create. "We need to be the ones making decisions about what's happening," said Kra. "If we aren't proactive about that, we will just be reactive, and we will not perhaps be happy with the outcome."

A common concern is that the AI labs' models are proprietary: Nobody outside the labs understands exactly how and on what data models are being trained, and ongoing access to the models can’t be guaranteed. There is also concern that the labs will increasingly dictate the direction of research, using their tools for a narrow set of problems and ignoring others, reshaping the field in terms of where both attention and money go. For some, the stakes run beyond mathematics. Naskręcki, who tests models for OpenAI, described unease at realizing that the same reasoning capabilities he helps benchmark on abstract algebra can be used for purposes he wants no part of. "I was actually very scared one day when I realized that, you know, if you help develop reasoning models, that can be easily used on the battlefield to just kill more efficiently," he said, comparing the position to that of the physicists on the Manhattan Project. 

To some extent, insulation from these pressures is a luxury pure mathematicians have enjoyed for centuries — the ability to pursue esoteric work without significant outside funding. "In pure mathematics, it's kind of shocking that your research might be guided by where your funding is coming from," said Vakil — a comfort long abandoned in the physical sciences, where expensive infrastructure has made industrial collaboration a fact of life.

Yet many researchers feel an imperative to work closely with the labs to understand what's coming. "We have to know what's there," said Vakil. "To not experiment, [to] not engage, seems not the right choice." Kra conditionally agreed. “One has to engage, but that doesn't mean one has to agree," she said. 

Longer term, said Michael Harris, a math professor at Columbia University and another co-author, there is hope that public computational infrastructure might emerge — a sort of CERN for AI, providing computing power to academics across disciplines so they do not have to rely on AI companies. 

A model for others

In some ways mathematics is distinct from other sciences: The majority of its results can be shown with certainty to be right or wrong, which is what makes it such an attractive target for labs looking to test their models and generate publicity. As a result, "the conversation in math is happening probably a couple of years before it happens in a lot of other areas," said Daniel Litt, a mathematician at the University of Toronto.

But AI will come for other academic disciplines as well. “Maybe in other fields they have substantially different methods, so they probably are going to use AI in a different way,” said Naskręcki, “but I think the same kind of issues arise.”

As that happens, the Leiden Declaration could be a template for how to respond. “It's the start of a conversation,” said Kra. “But I'm hoping that it will lead to concrete policies.”

 Image by Ian Lyman/Midjourney

AI is changing battlefield camouflage. In the war zones of the Russia-Ukraine conflict, conventional military camouflage is being replaced by bold black-and-white stripes in order to evade AI-powered drones. These systems use computer vision models trained on existing images to navigate and identify potential targets. It is possible, however, to trick AI by changing what the targets look like. The Economist reports that in the Russia-Ukraine conflict, Russian trucks and drones are being painted with bold stripes to deter AI detection, while some planes have been spotted with rows of tires on their wings to confuse aerial imaging systems. For now, most drones are controlled by humans even if they fly semi-autonomously, making these tricks less useful. But as systems become increasingly autonomous, this sort of adversarial camouflage will become more common. What ensues will be a cat-and-mouse game: AI systems will learn that sometimes trucks are striped, and new deceptions will be invented.

CRISPR is finding new ways to treat disease. Hear the term CRISPR and one probably thinks of editing DNA to repair or replace faulty genes. But a new generation of therapies takes a gentler approach. Rather than physically changing DNA, researchers are using CRISPR to control how genes are expressed without altering the underlying genetic code. The technique, known as epigenetic editing, relies on a modified version of the CRISPR protein Cas9. Known as "dead" Cas9, it can locate precise stretches of DNA but is engineered so that it does not cut them. Instead, it carries a protein that switches genes on or off. Because the DNA is not cut, the approach avoids some of the risks associated with conventional CRISPR, including accidental edits elsewhere in the genome or unwanted DNA rearrangements that can occur when a strand is cut. According to Nature, the resulting changes resulting from epigenetic editing appear long-lasting while remaining reversible if necessary. A small but growing group of startups is now betting on the approach. nChroma Bio, a biotech startup based in Boston, is using epigenetic editing to try to lower cholesterol by suppressing production of a certain protein and has reported promising results in monkeys; it is also working on advancing a treatment for hepatitis B. Tune Therapeutics is developing a similar hepatitis B treatment. Scribe Therapeutics, co-founded by CRISPR pioneer Jennifer Doudna, is, like nChroma, pursuing cholesterol therapies. And Epicrispr Biotechnologies is targeting a type of muscular dystrophy that affects the face and arms. The technology is still not entirely risk-free. Suppressing an incorrect gene — say, one involved in immune function or development — could have unintended consequences. Still, the approach is an evolution in CRISPR that could potentially make it a less risky treatment. 

Diamond could replace silicon in chips for use in harsh conditions. Silicon is the linchpin material in computer chips, and for many applications it's the ideal choice. But in some situations — handling large electrical currents, say, or in high-temperature conditions — silicon chips can become far less efficient or even stop working. Chemical and Engineering News reports that more than a dozen startups are now developing diamond semiconductors as an alternative. Pure diamond does not conduct electricity, but it can be engineered to include trace amounts of other elements — such as boron, nitrogen or phosphorus — that turn it into a semiconductor. Producing such engineered diamond was once prohibitively expensive, but the rise of lab-grown diamonds brought down the cost. The resulting diamond wafers can be turned into chips that work well where silicon falters. One example is the electronics that transfer high voltages from batteries to motors in electric cars. Silicon chips handle those voltages at roughly 90 percent efficiency, and many EV makers have switched to silicon carbide, which manages about 95 percent. Diamond chips, their developers claim, could reach 99 percent. Another potential market is power delivery in data centers, where silicon carbide is also increasingly used but where diamond promises to be more compact and efficient. The task now facing the diamond startups is bringing costs down enough so that diamonds’ better performance is worth paying for. 

What Happens if China Hacks the US Water Supply? I Went to a Secret War Game to Find Out, from Wired
3,000 words, or about 12 minutes

Imagine: A Chinese hacking group breaches thousands of water utilities around the United States, paralyzing their IT systems. In some cases the hackers have triggered physical destruction, bursting water mains. Soon, industrial facilities, hospitals and data centers are forced to shut down. Then water companies and other businesses, reeling from the effects, are calling their insurers to figure out what to do, at which point it becomes an insurance crisis, too. This Wired article reports on the experience of a few dozen insurance executives as they war-game this scenario. Should they prioritize their most valuable clients? Focus on minimizing harm to the public? Worry about whether the episode will bankrupt them? Obviously there's only so far you can push an exercise like this during an afternoon in a boardroom. But the point, according to its organizers — an insurance industry cybersecurity organization called CyberAcuView — was to shatter the industry's assumption that it could cope with such an event at all. The hope is that this realization would prompt insurers to use their policies to force clients to fix security holes that hackers could exploit. 

Why worms (and microbes) are catching on as a manure pollution solution, from MIT Technology Review
4,000 words, or about 17 minutes

Thanks to state regulation, California's dairy farms have become a test bed for cutting manure pollution. That matters more than you may think: Manure management accounts for about 1.6 percent of US greenhouse gas emissions, and in California — where the dairy industry produces roughly 45 percent of the state's methane — a 2016 law requires that levels be at 60 percent of where they were in 2013 by 2030. Well over a billion dollars has flowed to farms through state incentives, and most of the resulting emissions cuts have come from anaerobic digesters, which capture methane from slurry lagoons and sell it as fuel. But digesters make financial sense only for very large herds — a fact that has spurred interest in cheaper alternatives, including something called vermifiltration, the approach described in this story from MIT Technology Review. In this process, manure is separated into solids, which are composted, and liquids, which are filtered through beds of wood chips and crushed rock that are home to vast numbers of worms and microbes. The biofilter appears to sharply reduce nitrogen pollution, including the ammonia that ends up as nitrates in groundwater. The reduction in methane cuts is less clear: Studies backed by the technology's maker found significant reductions, but an independent study measured emissions as substantially higher than those produced by a conventional lagoon, and researchers say the question isn't settled. Even so, the technology demonstrates that there may be easy, untried solutions to tough problems. 

In Praise of Observational Evidence, from Asterisk
4,250 words, or about 17 minutes

Randomized controlled trials are broadly considered the gold standard for testing a hypothesis. The pharmaceutical industry is built around them. But they are often expensive and difficult to conduct, potentially preventing or slowing scientific progress. There is an alternative: using observational data. An enormous trove of information about people has been gathered through standard data such as electronic health records and routine surveys. This sort of data is often dismissed as not useful, as it can lead to biased or inaccurate findings. But this piece argues that sophisticated modern data science, combined with greater computing power, can be deployed to counteract those flaws. As a result, observational data can increasingly be used to emulate the approaches used in randomized controlled trials, making it possible to achieve high-grade insight where an RCT would be too costly, slow or unethical to run. RCTs aren't going anywhere, but the essay makes the case that we overvalue them, and that leaning on cheaper observational evidence could be useful in instances where a full trial was unlikely.