Notes from the Joint Mathematics Meetings
Quick thoughts written before leaving Seattle
Starbucks was a modest local establishment the last time I came to Seattle, and no one had heard of Jeff Bezos. As promised, AI was impossible to miss at the 2025 Joint Mathematics Meetings, but I didn’t hear anyone mention Bezos, although, in a February 2024 article in The Nation entitled “The Dirty Energy Fueling Amazon’s Data Gold Rush,” I learned that local environmental groups in northern Virginia
… estimate that the total capacity needs for approved-but-unbuilt data centers amount to 23.4 GW, the equivalent of 5.8 million homes (more than the total number of households in the state). Based on regional energy operator PJM’s forecasting, that will contribute to a doubling of the area’s peak energy demand by 2040 and will require a state grid with the same capacity as France’s.
Google tells a different story in its AI search function, which, when asked about Amazon and climate change, returns the sentence “Amazon Web Services (AWS) is a major consumer of energy, but the company has taken steps to reduce its impact,” and claims that
In 2023, AWS achieved its goal of matching 100% of its electricity consumption with renewable energy, seven years ahead of its original 2030 goal.
and
AWS intends to be water-positive by 2030, meaning it will return more water to communities than it uses.
Laura U. Marks addressed the environmental cost of AI in terms much more in line with the Nation article, in one of two sessions entitled Mathematics, AI, and the Social Context of Our Work, organized by Yaïm Cooper and Darren Byler. In fact, she understated the problem, to judge by Ezra Klein’s column in the January 12 New York Times. The future of AI, according to Anthropic’s Jack Clark,
will turn on power dominance: “Do you have access to enough electricity to power the data centers?”
“Dominance” here is an allusion to the competition between the U.S. and China. Klein continues:
A report from the Lawrence Berkeley National Laboratory estimates that U.S. data centers went from 1.9 percent of total electrical consumption in 2018 to 4.4 percent in 2023 and will consume 6.7 percent to 12 percent in 2028. Microsoft alone intends to spend $80 billion on A.I. data centers this year. This elephantine increase in the energy that will be needed not just by the United States but by every country seeking to deploy serious A.I. capabilities comes as the world is slipping further behind its climate goals and warming seems to be outpacing even our models.
The implications for the goal of keeping warming below 1.5° C are clear:
The addition of A.I.’s energy demand pushes back the finish line in a race we were already struggling to run, and the return of Trump, who intends to pull America back out of the climate accords, suggests that the world’s largest economy will cease to even try.
I gave a talk entitled “Mechanizing Mathematics: Who Decides?” at the first Social Context session, which provided a welcome opportunity to talk about questions that have largely been missing from discussions of AI among mathematicians: surveillance, military applications, labor practices, and the powerful interests behind the industrial development of AI, as well as the effects on the profession.
More such conversations will be needed before mathematicians will be able to judge whether Amazon’s claims (as transmitted by Google) or the figures reported by the Nation are more reaslistic. Apart from the two sessions on Social Context, I attended one on AI and mathematical publishing and (most of) a session called AI for the Working Mathematician. The former, where I presented an abridged version of this post, was a sober (but too brief) discussion of the opportunities and challenges posed by the emergence of AI as a major factor in mathematical publishing, with a clear emphasis on the challenges and unanswered questions.
The latter session was more about perceived opportunities. I asked two questions after a talk by Jason Rute, of IBM research, on “Applications of AI to formal theorem proving,” which ended with a series of predictions, all of which started with “AI will…” without specifying whether the AI would carry out these actions spontaneously or under human guidance. To my question about environmental cost, Rute mused aloud about whether the energy needed to train models will come from nuclear, solar, or fossil fuel plants; to my question about whether texts made public under a no commercial use license would be included in training data, he hesitated before deciding to “punt” on the question.
The JMM is over but the indispensable conversations about the implications of AI for mathematics are only beginning. When I asked outgoing AMS President Bryna Kra how it was decided to designate AI the “official theme”1 of this year’s JMM, she told me it had been her decision — answering the question I raised here — and explained that her purpose was not to celebrate AI unthinkingly, which has been the tendency in other meetings and in most of the press coverage, but rather to impress on the mathematical community the need to pay attention to the rapid diffusion of this technology and to anticipate its consequences for the profession. And indeed, the image of “mathematicians with their heads in the sand” came up frequently in conversation. Here is DALL-E 3’s attempt to represent this.
I found Kra’s explanation thoroughly convincing and am grateful for her leadership and for her acknowledgment that the issues around AI and mathematics are more contentious than the picture typically presented to outsiders. Feedback during the meeting convinced President Kra that the message was received, and that those in attendance are beginning to think seriously about what AI will mean for mathematics.
Meanwhile, several people came up to me at the conference center to tell me that they were regular readers of Silicon Reckoner, and to thank me for taking the time to put into words what many of them believe. I will judge the Seattle meeting to have been a success if, in the near future, there will be more venues, in addition to this one, where such readers can see their concerns reflected.
After publishing four posts in quick succession I will be taking a break to prepare the course I am teaching at Columbia this semester with the philosopher Justin Clarke-Doane. Readers will notice that students will be invited to ponder the implications of FrontierMath, which was widely publicized early last month. I met Elliot Glazer, the FrontierMath lead author, at one of the many JMM catered receptions. The article in Science reports that
experts think AI models will catch up to the new benchmark sooner or later—whether mathematicians like it or not.
As a public service, since nearly all the problems in the FrontierMath database are secret, I will be releasing my own list of 10 or so benchmark problems in a future post.
P.S. No one to whom I spoke in the AMS could tell me who chose the slogan “We decide our future.” Readers with information that can help solve this mystery are urged to share it in a comment, or to send me a private message. Anonymity guaranteed, as always.
Kra and others confirmed that 2025 is not the first time the JMM was organized around a specific theme. Climate was the theme in 2013, for instance (the only instance I’ve found, actually); but perhaps it was not explicitly called the official theme.