[No AI was used in composing this blog post.]
I’ve been trying to get back to the blog recently. But every time I plan to write a blog post, something happens that requires my attention and adds to the list of items I should be writing about.
I’ve decided that the best I can do, for now, is to summarize some of the key stories that have been on my radar. Because I’m covering so many topics so briefly, the chance of an error is pretty high; please let me know if you spot inaccuracies. Normally I try to be super-careful, but today I’m prioritizing conveying the sense of things rather than getting every detail right. Moreover, a lot of what follows is half-baked opinion; things are happening way too fast for me to be sure that my thinking is clear and correct, so you should interpret this post in that context.
The overarching news, aside from the 2026 physics Nobel Prize is what you’ve heard: artificial intelligence (AI) is transforming mathematics and high-energy theoretical physics, no question. That said, what is reported in the media about this is usually wrong in key details, and almost always wildly overstates what’s been accomplished. Nevertheless, the advances in AI models in the last couple of months are historic and transformative, and there is every reason to expect this to accelerate. For people like me, 2027 will surely be very different from autumn 2026, which in turn barely resembles spring 2026.
A first point: as many experts have noted recently, AI unchecked is extremely dangerous. This is especially true of artificial general intelligence (AGI) machines— machines that are capable of learning to do anything and everything we can, and can do it faster, better and tirelessly. AGI machines are, in my view, a path to slavery and death for a large fraction of humans — possibly a near-death sentence for our species. I personally believe that AGI development and usage should be banned, NOW. (A draft of a document laying out my argument is here; comments are welcome.)
However, specialized machines that are very good at a number of things, but very bad at others, are somewhat less dangerous. This includes tools that can do math and physics better than any human; they will be very useful, but dangerous if they can also command a robot army or shut down a banking system.
As for the new AI-led or AI-assisted results in math and science, they are coming thick and fast. Already in August and September, Anthropic claimed its models had made progress on the famous Riemann hypothesis, and had solved a particular 9-loop quantum field theory calculation. The latter is a rather esoteric achievement, even for most high-energy physicists, much less for the layperson. In a sense it’s a math problem, and not one that, by itself, leads to any fundamental new insights in physics that I’m aware of. It was also over-hyped in the media. But still, it is a significant accomplishment that surprised the experts.
The real flood began in early October. First, OpenAI claimed to have solved a key question concerning the notoriously difficult Navier-Stokes equation for how fluids flow. Initial wild rumors suggested that they had actualy solved this equation; had this been true, it would have meant that they understood turbulence, one of the oldest and greatest problems in all of science. The reality was far less; they had proven an important fact about the equation’s solutions. In a sense, this is one of the “Millenium Prize” problems (whose solution is worth a measly $1 million), but this is somewhat disputed, because the problem was solved under a certain assumption that some would say should not have been made. There are other controversies, involving whether certain mathematicians’ work was improperly used. Moreover, the work involves finding a counterexample, which is much easier than a proof of the absence of counterexamples. On top of that, the proof (even the version intended for humans) is excessively complicated, and challenging for a human to interpret. By itself, impressive for sure, but not nearly as much as the hype.
But then on October 6th, OpenAI dumped hundreds of papers on us, with results in mathematics and theoretical physics. Many of the results were interesting and a few seem important, including notably a further significant advance on the Riemann hypothesis. Unfortunately, most of the papers are long, technical and extremely difficult to read; they seem pretty useless. They may be correct, but there simply aren’t enough human experts in the world to verify them all and to translate them into shorter efficient summaries that can aid future human thinkers. Worse, imagine now that this is just the start: that hundreds of such papers, and soon tens of thousands, will be dropped on us every month. This will just bring the mathematically-oriented sciences to a halt, to no one’s real benefit. There has to be a better way to approach this.
Then on the 26th, OpenAI claimed three big new results in math, computer science and physics; the computer science result was also claimed, more or less simultaneously, by Anthropic. The physics result was widely and mistakenly reported as a solution of the key mystery of the theory of the strong nuclear force: why, in a theory that purportedly describes massless gluons, does the theory in the end exhibit no massless particles, with the gluons all confined into massive clumps (called “glueballs”)? (In jargon, this is called “generating a mass gap.”) Had OpenAI done that, it would have solved one of the central “Millenium Prize” problems. But despite the media hype, the reality was much less glamorous; OpenAI merely claims to have proved that the field theory that describes the strong nuclear force makes mathematical sense. That is a long way from proving that the strong nuclear force confines gluons (and by extension, quarks too, in a limited sense.) A proof of confinement would be worth a “WOW!”, and worth $1 million. A proof of existence, however, while definitely important, won’t affects most experts’ research; for them, it’s worth a raised eyebrow, and a “we knew that from experiment, but it’s nice to have a math proof”. And yet again, the proof is hundreds of pages long and apparently very hard for humans to interpret. It will take experts weeks or months to get through it and reduce it to a form that anyone else can handle — and only then are we likely to understand if the proof is actually useful for physics, rather than merely a technical feat that only a machine could love. There’s a pattern here…
But meanwhile, physicists are using these AI tools to try to solve problems that are of greater interest and perhaps greater importance than those that Anthropic and OpenAI have announced. So far these have tended to be pretty esoteric from the view of the general public, but one early October claim (by Nathanial Craig, who is a visitor at OpenAI and a professor at UC Santa Barbara) is that a technique may have been found, using OpenAI’s AI model “Astra”, to define quantum field theories like the Standard Model on a “lattice” — a grid in space and time. (This is notable because Craig, while an leading expert in particle physics, is not known as an expert in this subfield of “lattice quantum field theory”.) This technique might lead to a method for simulating these theories on computers, which has not previously been possible. In such theories, referred to as “chiral”, the masses of particles like the electron and the quarks can only become non-zero in the presence of a Higgs field (here’s why); and this makes numerical simulation extremely difficult, because (speaking very crudely here) massless electrons and quarks are normally very hard for computers to handle. If the paper is correct and leads to a practical simulation technique, it could prove very valuable for future developments in particle physics and elsewhere. Unfortunately the paper is over 100 pages, and despite Craig’s efforts to convert the work of ChatGPT-Astra into a human-readable form, it will be some time before the field digests it; I myself have no time to do so. And this is one of many papers that are appearing faster than I can even read the titles and descriptions, much less convey their essence to you.
In a totally different vein, my Harvard colleague Matt Schwartz used Anthropic’s AI models (“Claude”) to solve whole new classes of mathematical integrals that are important for particle physics. Then he used AI to look for problems in other areas of science (and beyond) where these same integrals also appear. Collaborating with experts in those areas, he has now used his tools in dozens of significant projects ranging across many disciplines. It’s quite a tour de force, and would have been impossible before mid-2026. You can read more about that here.
This whole process is generating a lot of political backlash. Through personal connections, some scientists are getting access to the AI tools and the AI companies in ways that create haves and have-nots, undoing decades of efforts at making physics a more quasi-democratic meritocracy. Increasingly it seems that producing the best science has less and less to do with talent, and more to do with wealth and power and one’s political standing and “who you know”. (A disturbingly secretive invite-only meeting at the KITP-UCSB last week did not help matters.) It’s the old boys club all over again, with billionaires running the show. I hope that this trend will not continue, as there will be a lot of bad blood if it does.
So high-energy theoretical physics is in the midst of a vast transformation. One way to say it: there will never be another Newton or Einstein, because even the smartest scientist will work with AI from now on, and AI will do most of the work, and probably will be producing some or most of the ideas. It’s not even clear there will be anything for human experts to do, in the long run, other than interpreting AI results for human consumption. But for real progress, AI models need to get a lot better at writing papers and simplifying their proofs, derivations and arguments. Any proof or calculation is worth a lot more when it is both correct and efficient, and scientists have historically gotten a lot of credit for simplifying past results that were nearly uninterpretable. This is partly why Julius Schwinger, the genius who could do Herculean calculations using elaborate mathematical methods, is not as widely known today as Richard Feynman, who did some of the same calculations using methods that were easy for everyone to understand and implement.
For all these developments in mathematical/theoretical science, we should remember that at heart, physics is an experimental science. AI can’t “solve the universe” by just thinking; it has to make predictions, and those predictions have to be tested in the physical world. Even if AI invents seven new theories of quantum gravity, or other ways of explaining the cosmos, it will not tell us which of these theories is relevant for our own universe. Without experiment, we can only know all possible worlds; we cannot know which one is our own. So experimental physicists certainly have not been replaced… not yet. To replace them will require great advances in robotics.
Finallly — old school — a particle theorist won a Nobel Prize for inventing a new way to look for neutrinos, and then leading the experiment that first carried it out. As far back as 1960, it was suggested by Moisey Markov (who died long ago and could not be considered for this year’s Nobel prize) that neutrinos could be searched for in a deep lake or ocean; on the rare occassions that a neutrino hits the nucleus of an atom in the water, the super-fast fragments from the collision emit Cerenkov light, which though faint can be observed in the abyssal darkness. However, there are other things that emit light in the ocean, and there are complications such as waves and storms (and salt, in oceans) that make building a stable detector challenging. In 1988 Francis Halzen had the clever idea of using ice instead of water — of melting deep holes into an ice sheet and putting light detectors into them. Ice has no waves and no critters emitting light; detectors in ice stay where you put them. The ice that lies below the south pole is both deep and pure, perfect for this effort. Unusually for a theorist of the time, Halzen chose to lead the experimental collaboration that built the first ice-based detectors, initially AMANDA and then the much larger IceCube. I wrote an article explaining IceCube here; it’s still current. So successful was this idea that it led to the birth of high-energy neutrino astronomy. Only very recently are deep water detectors catching up with IceCube, potentially making important discoveries of their own. Halzen’s Nobel-Prize winning idea (which, I’m told, occurred to him after a winter’s lunch in sight of the frozen lake that borders the University of Wisconsin, where he has spent his career) appears in a very short paper (with John Learned). It contains no math — much of it can be read by a layperson — and ends with a quaint diagram, reproduced below. The picture says it all. The prize isn’t merely for this paper though; it is also for shepherding the idea all the way to fruition, and ensuring that real discoveries about the universe were made with it.
As for me, I’ve been working on black holes with Harvard colleague Mark Riojas, and we’ve discovered some very interesting things (without using AI, I’m pleased to say!) I don’t have time to write about that now, but maybe I’ll get to it later. That’s in addition to some Large Hadron Collider physics, mostly on this and related subjects, and a public outreach workshop at Canada’s Perimeter Institute.
Such is the news for today. If I sound frazzled, it’s because I am. It’s a time of great chaos. I suppose we’ll muddle our way through it somehow.
