Concern about Industrialization of Theoretical Physics by AI

This longg post regards theoretical high energy physics (hep-th). In particular, I'm talking about string theory, supersymmetric theories, supergravity (it also applies to the field of mathematical physics). Basically the part of hep-th largely disconnected from experiments.

This subfield of hepth has maybe ten thousand faculties and postdocs worldwide. The usual research process is talking to colleagues + going to conferences -> getting ideas -> doing computations -> putting papers on arXiv -> publishing in journals. Each cycle takes about 1 to 2 years (faster if you're being scooped).

It takes a year for a new PhD student and months for a newly hired postdoc to become really useful in doing computations and collaborating with you in your very specific subfield.

A week ago, I gave GPT-Astra all my papers and previous computations. It ate 90% of my weekly token allowance (on the $500 monthly plan), but by the end it could reproduce almost all my results and do new computations (i thought couldnt be done without some crazy cluster) in the sub-sub-sub field that I am THE expert on.

Addicted with Astra, two days ago I used Astra and Sol 6.1 to invade investigate other subfields. It produced results that I independently verified, including computations that colleagues who are world experts in those subjects had struggled with for a decade. (Let me clarify, othes had obtained the results through other very indirect methods years ago, but this is the first time this computation was done using the general prescription from 2005, something the community always wanted to do)

Six months ago, when I saw papers almost entirely written by AI in our field, they were complete garbage with hallucinated references and nonsensical results it proudly claims. Lol things have changed. Just last week, a professor in our field published a paper explicitly saying that everything from the original idea to the final writeup was done by AI, with humans checking the results before putting it on arXiv. And the results were actually really good. The paper was also fucking well written.

Now let's talk about journals and referees.

I've published dozens of papers in (journal of high energy physics) JHEP. Many referee reports are honestly pretty useless. They are almost always: one paragraph summarizing your paper (basically copying your abstract), another paragraph saying it's interesting and recommending acceptance with minor changes, followed by a whole page, in point-form, of clarification to certain sentences and grammatical corrections. Sometimes there are critical typos that I found out only years after its published.

But when I asked AI to referee my papers before submission, the reports were ridiculously detailed. It reproduced my computations, investigated the future directions I proposed in the conclusion (and actually did some of those computations and told me they dont work), and found obscure missing references from old pure math papers that only had two citations over 40yrs that i've never heard of and wasn't even available on arXiv.

So it seems to me we are already at a point where Astra level AI can come up with ideas, do computations, write papers and referee papers better than the average researcher. I'm not talking about AI papers better than the best of the best of the best in our field. But the average researcher? I think we are already there.

And this is TODAY. Imagine two years from now, with models two or three generations ahead, running on Rubin chips or Feynman chips or whatever.

Imagine this.

An AI startup company like FirstPrinciples in 2028, running tens of thousands of AI agents with vastly more tokens than any individual faculty member or university research group can afford. These agents continuously come up with and investigate hundreds of research problems simultaneously, producing tens of thousands of pages of research, neatly written in LaTeX every hour.

Then another set of agents, ideally using different models like Claude, Kimi, Gemini etc, independently verify everything with given tasks of poking holes in the papers in anyway they can. The writing agents and referee agents go back and forth, correct errors, and eventually distill the most solid and consequential results.

At the end of each day, the system outputs maybe 20 PRL or PRX level papers, each around 6 pages long. A small group of highly paid theoretical physicists who ditched academia and hired by the company reads them and gives the final stamp of approval.

Every month, the company sends its ten most important papers to leading theorists at IAS, Stanford, Harvard, etc for their opinions, similar to what OpenAI is already doing today. Then do a news report on the results.

And honestly, in few years, if the AI results get the human stamp of approval nearly 100% of the time, why would you even need that human approval anymore? Just a 24/7 paper mill producing ridiculously high ass quality theoretical physics papers.

Now people will say, who is going to fund this? Hepth doesn't make money.

Well, all you need is someone like Elon Musk or Jared Kaplan (who used to be a string theorist and is now one of the founders of Anthropic) waking up one morning and deciding to donate $300 million to a few autonomous hep-th AI startups. These guys already donate hundreds of millions to universities. Just instead of having a building named after them, they can insist, since AI cant be author, that they are the only human listed as author in all the papers. Some of these papers may be as relevant as Feynman's paper on pathintegral! hundreds of mills is nothing for these people so of course there is incentive to do so! And they will certainly feel pride looking at their name on the paper because the result is indeed due to computing power funded by their donations.

I truly fear that within a few years, a handful of well funded AI research companies could produce order of magnitude more great papers in hepth research than the entire existing community.

And I don't think our field is remotely prepared for this. But also is this a good thing for hep-th as a field even if its not for the humans in the community? Maybe... I dont know.

What happens to PhD students? Postdocs? Faculty hiring? Journals? What is even the point of spending years training someone to do computations if an AI can learn your entire research program in a few days? As faculties is it in good conscious to get new PhD students knowing the possibility the field no longer exist in a human friendly way by the time they graduate? Am I missing something here? Because I feel like people in our field are seriously underestimating how quickly this could happen! And especially since we dont have the bottleneck of waiting on experimental results.

TLDR: GPT-Astra can exercise every part of the research cycle as good as a researcher in high-energy theoretical physics. If start-up companies automate this process and throw money at it, I don't see how the community can compete. And all of these may happen in two or three years.