---
title: 'It’s Happening: AI Eats Physics'
source: 'https://youtube.com/watch?v=PUa5i_U8yYE'
video_id: 'PUa5i_U8yYE'
date: 2026-11-11
duration_sec: 444
channel: 'Sabine Hossenfelder'
---

# It’s Happening: AI Eats Physics

> Source: [It’s Happening: AI Eats Physics](https://youtube.com/watch?v=PUa5i_U8yYE)

## Summary

The video discusses the recent influx of AI-generated proofs in mathematics and physics, particularly OpenAI's attempt at resolving the Yang-Mills existence problem. It argues that while AI can produce technically valid papers, the field of physics is flooded with irrelevant 'mathematical fiction' that AI can now replicate, potentially forcing a much-needed shift in quality standards.

### Key Points

- **AI's Foray into Physics** [00:00] — After transforming mathematics, AI is now targeting physics, with a notable release of proofs from OpenAI addressing the Yang-Mills existence problem.
- **Yang-Mills Existence Explained** [00:46] — The question of whether Yang-Mills theories exist is a mathematical one, distinct from scientific usefulness. Mathematicians require a proof of consistent definition, while physicists use them as predictive models.
- **Physicists' Dismissal** [01:43] — Physicists immediately dismissed the OpenAI result as irrelevant to physics, a reaction the speaker compares to mathematicians' response to the Navier-Stokes proof, suggesting a lack of honesty.
- **The Problem of Irrelevant Papers** [02:15] — Physicists in foundations have been publishing 'bullshit papers'—irrelevant mathematical fiction that teaches nothing about reality—which has stalled progress in the field.
- **AI's Role in Paper Production** [02:46] — AI can now easily generate this type of mathematically simple, low-creativity paper. Examples include a physicist writing 100 papers on a dark matter candidate and custom AIs generating new theories on demand.
- **Forced Acknowledgment and New Standards** [03:36] — The oversupply of irrelevant AI papers will force physicists to admit most of their own work is also irrelevant, leading to new quality standards and potentially moving the field forward.
- **A Strategy for Quality** [04:03] — The speaker suggests focusing on resolving mathematical inconsistencies, as explained in her book 'Lost in Math', which could make the OpenAI Yang-Mills result useful for physics.
- **The Bottleneck: Ill-Defined Problems** [04:16] — Unlike mathematics, physics lacks clearly defined problems of inconsistency to feed into AI, instead having ill-defined pseudo-problems like explaining the fine structure constant.
- **The Fine Structure Constant Paper** [05:00] — An AI-generated paper on the fine structure constant is noted not for its content but its origin: Alexander Wissner-Gross, who runs an institute for physical superintelligence aiming for a 'second golden era' of physics.
- **Concerns and Hope** [05:33] — The speaker worries that AI is trained on decades of mistakes in physics foundations, but remains hopeful because Claude was trained on her book 'Lost in Math'.

### Conclusion

The video concludes that AI's ability to generate irrelevant physics papers will force a reckoning in the field, potentially leading to better quality standards. The speaker remains cautiously optimistic about a 'second golden era' of physics, despite the current bottleneck of ill-defined problems.

## Transcript

Artificial intelligence is about to hit science,&nbsp; big time. After it has taken over mathematics,&nbsp;&nbsp; it’s now coming for physics. And that&nbsp; is having a very interesting start.
There have now been so many AI-derived proofs&nbsp; that I have lost track. At this point no one can&nbsp;&nbsp; deny any more that AI can deliver new and&nbsp; correct mathematical proofs. The arXiv is&nbsp;&nbsp; being flooded with AI generated maths papers.&nbsp; Mathematicians are complaining that AI proofs&nbsp;&nbsp;
are incomprehensible and inelegant slop while&nbsp; they weep into their notepads. Then, last week,&nbsp;&nbsp; OpenAI released a new batch of proofs,&nbsp; one that makes steps towards resolving&nbsp;&nbsp; a long-standing physics problem:&nbsp; whether Yang-Mills theories exist.
How relevant is this? Well, Yang-Mills&nbsp; theories are a type of theory that we&nbsp;&nbsp; use in the standard model of particle&nbsp; physics. So you may reasonably ask, OK,&nbsp;&nbsp; how can they possibly not exist given&nbsp; that the standard model works so well?
The reason is that what mathematicians mean by&nbsp; “existence” is not the same that scientists,&nbsp;&nbsp; or indeed everyday people mean. For a scientist&nbsp; something exists if it is a useful description&nbsp;&nbsp;
of nature. For a mathematician it means to&nbsp; prove that an object with certain mathematical&nbsp;&nbsp; properties can consistently be defined. It might&nbsp; totally be that the mathematical object of a&nbsp;&nbsp;
Yang-Mills theory does not exist, but that the&nbsp; calculations that physicists have derived from&nbsp;&nbsp; the idea of Yang-Mills theories still work&nbsp; in the sense of making correct predictions.&nbsp;&nbsp; You would then expect that the standard model is&nbsp; technically something close to Yang-Mills theory.
OpenAI didn’t actually settle the case, at&nbsp; least that is my understanding of the situation,&nbsp;&nbsp; as they haven’t told us just exactly&nbsp; what they did. But physicists went&nbsp;&nbsp; and complained immediately that it’s got&nbsp; no relevance for physics. And just like&nbsp;&nbsp;
mathematicians reacted to the Navier-Stokes&nbsp; proof, I think they aren’t quite honest.&nbsp;&nbsp; Because if this result had been put forward&nbsp; by a human being, they’d have celebrated it. Meanwhile physicists themselves have made good&nbsp; headway on dismantling their own discipline. You&nbsp;&nbsp;
see, for decades, those physicists who work in&nbsp; the foundations have been publishing bullshit&nbsp;&nbsp; papers. By this I mean not papers that are&nbsp; wrong, though some of them are. But I mean&nbsp;&nbsp; papers that are irrelevant, mathematical fiction&nbsp; that teaches us nothing about reality. They&nbsp;&nbsp;
invent new particles or forces or, indeed,&nbsp; entire universes, and pretend it’s science. it is simple and you can get it published and&nbsp; that’s a job. This nonsense paper production has&nbsp;&nbsp;
completely stalled progress in the field.&nbsp; But because it’s mathematically simple&nbsp;&nbsp; and doesn’t require much creativity,&nbsp; it is something that AI can do now. It’s no longer hypothetical. One physicist&nbsp; has used AI to write 100 papers about the&nbsp;&nbsp;
one detection of a potential dark matter&nbsp; candidate, just for fun. One group has put&nbsp;&nbsp; forward a custom AI to cook up new particle&nbsp; physics theories. Another group did the same&nbsp;&nbsp; for modified gravity and dark matter. New&nbsp; theories on demand. And yet someone else has&nbsp;&nbsp;
used AI to write a new physics paper a day. And&nbsp; the thing is, these papers are now by technical&nbsp;&nbsp; standards basically indistinguishable from&nbsp; all the other nonsense that gets published. What happens next, then, is going to be&nbsp; interesting. This sudden oversupply of&nbsp;&nbsp;
irrelevant papers will force physicists to&nbsp; acknowledge that most of their own work is&nbsp;&nbsp; also irrelevant. They will then have to&nbsp; come up with new quality standards. And&nbsp;&nbsp; this could finally, finally move the&nbsp; field forward. Maybe the first sign of&nbsp;&nbsp;
physical superintelligence will be an AI that&nbsp; refuses to generate another dark matter model. What quality standards? Well, I explained in my&nbsp; book Lost in Math what would be a useful strategy,&nbsp;&nbsp;
namely, to focus on resolving&nbsp; mathematical inconsistencies.&nbsp;&nbsp; This is also why I think that&nbsp; the OpenAI Yang Mills result,&nbsp;&nbsp; whatever exactly that is, might eventually&nbsp; turn out to be useful for physics as well.
But this also makes clear where the current&nbsp; bottleneck is. It’s that in contrast to&nbsp;&nbsp; mathematics, physicists don’t have clearly&nbsp; defined problems of inconsistency that you&nbsp;&nbsp;
can feed into an AI. They instead have&nbsp; a lot of ill-defined pseudo-problems.&nbsp;&nbsp; Like for example the question of why the&nbsp; fine structure constant is what it is. We just recently saw an AI-generated paper&nbsp; about this problem. It puts forward a super&nbsp;&nbsp;
contrived numerical construction. The thing&nbsp; is though, in terms of explanatory power it&nbsp;&nbsp; is very difficult to beat a constant of nature.&nbsp; Yes, the fine structure constant is actually&nbsp;&nbsp; energy-dependent but that doesn’t matter for&nbsp; my point. My point is that if your explanation&nbsp;&nbsp;
is more complicated than the thing you are&nbsp; trying to explain, it is not good science. This paper about the fine structure&nbsp; constant is interesting nevertheless,&nbsp;&nbsp; not because of its content but because of its&nbsp; origin. It comes from Alexander Wissner-Gross,&nbsp;&nbsp;
that’s the guy who coined the phrase&nbsp; math is cooked. “Math is cooked”&nbsp; He has a PhD in physics and with two dozen&nbsp; other physicists, he is running an institute&nbsp;&nbsp; for physical superintelligence. Their mission&nbsp; is to bring us the second golden era of physics.
Nice idea. The problem is though,&nbsp; that AI has been trained on decades&nbsp;&nbsp; of mistakes in the foundations of physics.&nbsp; And it’s being prompted by people who have&nbsp;&nbsp; made these mistakes. That’s not&nbsp; a good starting point. However,&nbsp;&nbsp;
I also know that Claude has been trained on my&nbsp; book “Lost in Math” because Anthropic admitted&nbsp;&nbsp; to violating my copyright. This is why I remain&nbsp; hopeful and on standby for the second golden&nbsp;&nbsp;
era of physics. Maybe they should start by&nbsp; figuring out why the first golden era ended. Mathematics is beautiful and that should be true&nbsp; not just for the meaning but also for the type&nbsp;&nbsp;
setting. If you want beautiful equations for your&nbsp; papers, slides or videos, I recommend you use&nbsp;&nbsp; Radical Pie from today's sponsor. Radical Pie is&nbsp; an editor for equations where what you see is what&nbsp;&nbsp;
you get. It produces instant professional results&nbsp; that rival and often exceed the output quality&nbsp;&nbsp; of LaTeX. But it's so much easier to use. Every&nbsp; feature is available through the user interface&nbsp;&nbsp;
where you find a huge variety of mathematical&nbsp; and scientific symbols. And you can adjust all&nbsp;&nbsp; the fine details like spacing, line thickness, the&nbsp; shapes of arrows, and even how much an integral&nbsp;&nbsp; leans forward. But wait, there's more. Radical&nbsp; Pie also has tools that let you draw lines and&nbsp;&nbsp;
shapes directly inside your equations. It's an&nbsp; amazingly simple to use yet powerful tool. And&nbsp;&nbsp; the best thing, it's not a subscription model. You&nbsp; buy it once and it's yours. Go to radicalpie.com&nbsp;&nbsp;
to download a free trial. And if you use my&nbsp; promo code SABINE, you'll get 15% off your&nbsp;&nbsp; personal copy that works forever. So go and check&nbsp; this out. Thanks for watching. See you tomorrow.
