In a previous posting, I’d noted that the number of arXiv hep-th preprints/month was significantly higher recently, clearly due to LLM use, but that I hadn’t heard of much in the way of significant new results coming from this. Around the same time, Matt von Hippel issued a challenge to the LLM agents:
So if AI companies want to impress people like me (or scare us, for that matter), then they need to tackle my old field. Show that an AI can take the kinds of computer resources an academic has access to, and solve one of the scattering amplitudes field’s big outstanding problems. Show that a computational limit everyone expected to be a problem doesn’t actually matter. Give us N=8 supergravity to seven loops, or N=4 super Yang-Mills to nine loops.
People at Anthropic took up the challenge, and today announced that Yes, Claude can do Nine Loops. Supposedly this was a “one-shot”:
after asking Claude which problem it was most likely to be able to tackle, they gave it a simple prompt:
“The problem is to compute the Six-particle (hexagon) amplitude in planar N=4 SYM at nine loops.”
From there, they just kept telling it to keep going, with comments like:
“I’m going to sleep and won’t be available for another several hours. Keep working on this until I tell you to stop. Give me updates every 4-6 hours.”
After using one to two thousand dollars worth of Claude time, they got a result. This was sent on September 1 to Lance Dixon, who had techniques that allowed him to validate the result.
The computation of amplitudes like this has become a significant part of hep-th in recent years, heavily promoted by Nima Arkani-Hamed. There’s no practical application of knowing things like this nine loop amplitude, but Arkani-Hamed always has emphasized that the motivation is to use such calculations to get inspiration for a very different fundamental definition of the theory in terms of new geometrical objects. I have no idea whether this nine loop result is helpful for that. It appears that Claude did not come up with a new method, just used known calculational methods. This is an extremely complex sort of calculation.
It will be interesting to see how these new tools change the way those in the large community of theorists doing such calculations pursue their work and publish their results. Will the new AI tools help them come up with new methods for computing amplitudes and new insights into their structure, or will they just make them obsolete, replaceable by any button-presser with a couple thousand dollars and no expertise?
At the other end of hep-th, there’s the even larger community of theorists who, inspired by AdS/CFT, believe the right way to understand quantum gravity is in some sort of quantum information/holographic terms (“it from qubit” is the slogan). Writers at Quanta magazine have always been quite enthusiastic about this direction in hep-th, and today Charlie Wood has a piece about it, entitled Gravity Seems Holographic. What Does That Mean for Reality?. Wood explains the background and latest thinking in this research direction, concluding:
At this point, I subscribe to a more extreme variation of the it from qubit interpretation — mostly just following the rumble of the stampede. I’d bet that the qubits are the real things, but that they don’t live on anything as familiar as a flat screen.
Physicists have tried to stretch the AdS/CFT correspondence to fit de Sitter space — which has no obvious screen — for decades, with limited success. In recent years, they’ve started to get more creative. Susskind and other teams have made progress (opens a new tab) on holographic de Sitter models (opens a new tab) that differ radically from AdS/CFT. Instead of squashing a volume into an area, these universes seem to cram all the dimensions into a lone quantum point. I imagine a bunch of quantum pixels all coexisting in one spot, rather than spreading across a screen. That might be hard to visualize, but we’re already accepting the idea of dropping one dimension of space. Why should tossing the others be so different?
Balasubramanian suspects that even this kind of radical model doesn’t go far enough. Einstein’s theory fused space with time, and so if the three dimensions of space emerge from a spaceless point, then time should emerge from something timeless. Somehow, we and everything we experience exist within an unblinking dot of no size.
I don’t think we’re there yet, but if Claude does become super-intelligent, way beyond our best theorists, we should be able to soon ask: “Claude, is this bullshit or not?” I’m not super-intelligent, but bet that I already know the answer…
Update: More from Matt von Hippel.
Update: Claude has made a new contribution to theoretical physics, finding a computational error in an Appendix of a well-known paper from the 1980s.
I think that the link in the update should point to https://4gravitons.com/2026/09/25/it-got-to-my-field/
Andrzej Daszkiewicz,
Thanks! Fixed.
The current AI models are extremely bad at evaluating the significance of a paper, other than going by what other people have said about it. They will default on criticising irrelevant minor details, rather than pointing out the obvious elephants in the room (eg: this entire research direction is useless, the answer to the question is irrelevant either way, etc). The reason is, of course, that no such evaluation exists anywhere in the published literature either. That, I think, is the real problem. Maybe AI will one day solve it, but it will be a while.
Agreed that current models cannot identify the significance of papers (and less generally the significance of physical arguments of which the papers are composed).
I am curious to know how this might relate to the current or future role of AI in theoretical physics; e.g,. if AI meets a barrier in theoretical physics whether that potentially says something important about both theoretical physics and AI itself.
1. Matt von Hippel’s challenge seems to involve a technically challenging calculation that specifically referenced computational requirements. That seems close to AI performance / successes in preferred domains (natural language processing, software engineering, and formal proofs in mathematics) we’ve seen recently.
2. Theoretical physics, in comparison, has two features that appear distinct to the field: Physical intuition and the relationship to experimental physics.
3. With respect to experimental physics, AI’s roots in data science suggest a potential close fit for AI in physics; as evidenced by ML / AI techniques in high energy, condensed matter, astrophysics etc.
4. However, in terms of “physical intuition” and its role in theoretical physics I am very curious to see how this plays out. For example, the success of AI in other fields (SwE, mathematics, Matt von Hippel’s challenge) seem rooted in the algorithmic nature of the problems; as opposed to problems that require an evaluation of interpretation, significance and “taste”; each of which contribute to “physical intuition”. This argument is similar to the question in mathematics about the role of AI in theory building and how that is distinct to technical problem-solving. This argument also seems to relate to the use of language models to process physical arguments; as required in theoretical physics and distinctly to NLP, SwE, scientific computing and formal mathematical proof.
Part of my curiosity in this matter relates to how theoretical physics is distinct to both mathematical physics and mathematics.
Potentially, as we learn more about where and how AI is more and less successful in these particular fields we will learn more about these fields themselves (and not just strengths / limitations of AI).
In my own field (astrophysics), AIs have probably reached the stage where they can produce something publishable quite easily – take this existing dataset, apply established analysis techniques to it, write up the findings in a 12 page paper. This is ‘publishable’ in the sense that you can get it accepted in one of the more reputable journals, and there are usually multiple examples of this kind of paper on the astro-ph arxiv on any given day. Whether any of it is scientifically useful is a separate question. I can see the benefit of using AI to speed up writing papers presenting new data (‘we ran these simulations/observed these objects, and this is what they look like’), saving human time for working out the broader scientific implications, but I suspect we’ll just end up with a lot more papers saying nothing interesting.
I’m struggling to see what the big deal is here. We’ve learned that LLMs have gotten good enough to use conventional methods to do a highly complex calculation with minimal supervision. If they’re not already better than people at this sort of thing, they soon will be.
I would have thought such a development would be regarded as entirely predictable and welcome, instead of there being surprise and angst. Relying on humans to engage in such herculean tasks seems about as sensible to me as expecting every neighborhood postal carrier to be an Olympic marathon runner. That would be impressive, but aside from admiration of ability for its own sake, I don’t see the point in that. Humans generally are not as good as machines (with the possible exception of the occasional von Neumann-level supergenius) at loads of brute-force calculating tasks. Hence we make use of those machines. This latest news seems like the latest iteration in our long history of replacing human calculators with artificial ones. Those calculators are now out of a job, and that’s regrettable, but hardly anything new, nor anything that’s likely to stop happening.
And since there was nothing new about this development than the passing of a certain technological milestone on that long road, what’s the larger significance? I can’t see any. It certainly doesn’t seem to rise to anywhere near the level of concern as recent stories about rogue agents. It humbles we humans, but we always had too high of an opinion of ourselves anyway.
LMMI,
I don’t think there’s much angst about LLMs doing complicated calculations in theoretical physics. Those doing such calculations will adapt, can just do more calculations, or the same ones with less hassle.
This news isn’t surprising, it’s just a good indication of where exactly the technology now stands. The interesting question is still open: whether these things can come up with significant new insights, or just be able to do the same not particularly significant things that theorists have been doing for a long time now.
“The interesting question is still open: whether these things can come up with significant new insights, or just be able to do the same not particularly significant things that theorists have been doing for a long time now.”
That could be where the difference between math and theoretical physics lies. One reason why LLMs are so good at math is that there is a lot of high-quality training data available on the internet. But according to some experts (including some well known bloggers), the training data in theoretical physics is not so great…
Pascal,
Yes, I’d argue that the reason we’re seeing a very different story in math than in hep-th is that math has been a much healthier subject the past few decades, with a continuing influx of good new ideas. hep-th suffers from the opposite problem: for a long time the new ideas that were getting attention were not good ones.
There have been real advances in the business of computing amplitudes, which is why AI agents could have some success there exploiting these.
On a separate front, hep-ph, see this blog entry about using AI:
https://rezaebadi.com/blog/lz-simulated-response/
The hep-ph long-term problem has always been a very different one than the hep-th problem. There the problem is how few solid experimental results conflict with the Standard Model. Kind of like what mathematicians are worried about: if there are no significant open problems, the field will wither away.
You may find this interesting, as it documents another way that AI changes how we do science. Jesse Thaler describes preparing and giving an AI-generated plenary talk at the LHCP conference in Paris (https://indico.cern.ch/event/1537360/)
“Before the physics, I should disclose how the slides I presented at LHCP were made: Claude Opus 4.7 produced them under my direction, from the selection of papers and the reading to the structured notes, the drafting, and the fact-checking. The process took roughly 72 hours of compute and went through multiple drafting passes, independent reviewer agents, and a verification pass
over the numeric claims. The talk was mine; the work was the AI’s.”
https://arxiv.org/pdf/2609.32874
Kinda weird and I have very mixed feelings. It’s novel and curious and provocative, but fundamentally I see talks as communication of ideas between humans. AI talks are to me an unwanted and unnecessary encroachment.
Hi Peter,
Have you tried using AI to research your own physics ideas?
Justin Glick,
I’ve been using AI a small amount, basically to ask for explanations or simple mathematical facts in cases where I’m trying to use mathematics that I’m not expert in. Sometimes this has been helpful, sometimes not.
The main problem is that mostly the process of trying to make progress on what I’m doing is one of starting from being completely confused about something, and getting to the point where things are clear. In the confused state, I don’t know what the right question to ask is. If I knew the right question, the answer would probably be clear. This same problem is the one that in the past has always been there when I’ve had the opportunity to talk with people far more expert than me on some techniques I’m trying to use.
From talking to other people using AI, I’m often told that I’m probably not doing it right. One should not just ask questions like I’m doing, but should set up agents in a certain way, etc. Maybe this is right, but my reaction right now is that, if it’s true that the capabilities of these things are improving dramatically, I don’t want to spend time figuring out how to get the current version to work well, happy to wait a bit.
From what Tsimerman said in the panel discussion mentioned here a while ago, April is when these things will overwhelm us. I can wait til then…
https://blog.arxiv.org/2026/10/01/updated-rate-limit-policy/
> In September of 2016, arXiv received 9,869 submissions. In September of 2024, arXiv received 20,569 submissions. This September, arXiv received 40,363 submissions, which in turn generated almost 9,000 support tickets for arXiv staff and moderators.
> arXiv now limits submitters to up to two submissions per calendar month, with a limit of three total active submissions at any given time.
According to
https://bigthink.com/starts-with-a-bang/ai-first-breakthrough-theoretical-physics/
some AI has calculated some kind of “best yet” amplitude for a QFT problem.
It might turn out that AI will turn out to be primarily a “super calculator”, producing numbers based on well established methods that are too involved for any human to undertake. We shall see.
Bryan,
That’s interesting. Unfortunately the arXiv doesn’t seem to have figured out how to deal with the onslaught of slop. This latest measure may stop some egregious cases, but the underlying problem remains.
I’m starting to get convinced that my initial reaction to this was right: training on the recent hep-th literature in its sad state produces AI capable just of producing stuff even worse. Opening the arXiv to vast amounts of this stuff will be a disaster. As of this year, the AI is being trained on an increasingly slop-ridden source. In hep=th we’re more likely to see model collapse and worse slop than to see AI that has a clue and figures out how to do better.
Raoul Ohio,
That article is about the same result discussed in this posting. I disagree that going from eight to nine loops using the same methods as before is a “breakthrough”.
Peter, yes I think there are 2 related problems here and arXiv is going after the easier one. The first problem is the onslaught of submissions, and arXiv’s solution should slow it down considerably by putting these submission limits in place.
The second, and more interesting problem, is identifying high quality papers. Just speculating here, but there are 2 required parts: one is a more sophisticated search ability so researchers can find papers that might be of interest to them in this massive pile of submissions. Of course this search would likely need to rely on AI agents to help surface the best papers for each researcher.
The second part is grading the identified papers. Currently this is the one step that can only be done by humans. After enough grading by humans, then of course the labels would be used by AI to train to writer “higher quality” papers, and the cycle repeats.
But in order for this system to work there needs to be some cost of submitting, and registered accounts with hard limits is one way to do that.
Any comments on the Buckmaster talk and the post-talk discussion today?
In related news, Schwartz published https://bootloops.ai/. It was criticized quite harshly at our physics department lunch chatter today for being AI slop… Any takes?