Physicists Tap AI for Research as ArXiv Theory Submissions Jump 30%

Sabine Hossenfelder

While automated theorem proving plunges mathematicians into an existential crisis, theoretical physicists are embracing artificial intelligence to accelerate peer-reviewed calculations.

AI-assisted summary of AI Has Eaten Maths. Physics Is Next — original source. Check the original for full context.

Source published: 2026-09-01

Key ideas

Artificial intelligence models have achieved major mathematical breakthroughs, uncovering a counterexample to a 1939 conjecture and making progress on the Riemann zeta function. (0:00)
The rapid automation of mathematics has triggered an existential crisis among early-career mathematicians, sparking calls for moratoriums on AI-generated proofs. (0:00)
Theoretical physics is experiencing a parallel shift, with arXiv high-energy physics theory submissions jumping by approximately 30% as researchers feed backlogs of ideas into LLMs. (2:17)
Unlike mathematicians who debate their discipline's survival, physicists are pragmatically adopting AI tools to solve problems ranging from quantum entanglement to fusion stellarator design. (2:17)

Detailed notes

## Background This video was published on YouTube by Sabine Hossenfelder on September 1, 2026. According to the page description, artificial intelligence has rapidly disrupted mathematics by solving long-standing open problems, leaving mathematicians in turmoil, while theoretical physics is quietly poised to be transformed next.

## TL;DR Sabine Hossenfelder examines how modern large language models, including Claude and ChatGPT, have rapidly tackled major mathematical milestones such as counterexamples to decades-old conjectures and advances on the Riemann zeta function. These automated successes have triggered widespread existential dread, anger, and calls for moratoria among early-career mathematicians. In contrast, physicists are pragmatically adopting AI to derive complex proofs, refute classical electrodynamic rules, model experimental optics, and optimize fusion stellarator designs. Hossenfelder highlights that while mathematicians fear for the philosophical future of their discipline, pragmatic physicists are simply delighted to automate tedious calculations and accelerate their publications.

## Atmosphere & style The video carries Sabine Hossenfelder's trademark dry, sardonic wit paired with an analytical, grounded realism. She highlights the dramatic contrast between the apocalyptic despair of pure mathematicians and the unapologetic, deadline-driven pragmatism of physicists, delivering each observation with deadpan comedic timing. The delivery feels brisk, informative, and sharp, capturing the unsettling speed of technological change without losing a cool, cynical sense of humor.

## Bulletin - [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) Artificial intelligence models have achieved major mathematical breakthroughs, uncovering a counterexample to a 1939 conjecture and making progress on the Riemann zeta function. - [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) The rapid automation of mathematics has triggered an existential crisis among early-career mathematicians, sparking calls for moratoriums on AI-generated proofs. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) Theoretical physics is experiencing a parallel shift, with arXiv high-energy physics theory submissions jumping by approximately 30% as researchers feed backlogs of ideas into LLMs. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) Unlike mathematicians who debate their discipline's survival, physicists are pragmatically adopting AI tools to solve problems ranging from quantum entanglement to fusion stellarator design.

## Key takeaways - [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) Claude assisted researchers in discovering a counterexample to the Jacobian conjecture, an open problem dating back to 1939. - [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) OpenAI announced ten major results from its unreleased Astra model that reportedly solve or advance open mathematical problems. - [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) Anthropic reported that Claude achieved significant progress on the Riemann zeta function toward resolving the Riemann hypothesis. - [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) The broader mathematical community increasingly acknowledges that large language models can perform rigorous, serious mathematical research. - [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) Graduate students and early-career researchers are facing profound anxiety and anger over the automation of math, with some demanding bans on LLM-assisted proofs. - [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) Daniel Litt warned that AI could overwhelm academia with automated, verified papers that no human will read, care about, or use. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) Theoretical physicists are feeding dormant research concepts into LLMs to generate calculations suitable for peer-reviewed journals. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) High-energy physics theory (hep-th) submissions on arXiv spiked roughly 30% in May and June compared to previous years. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) Gerardo Adesso used ChatGPT to develop the core proof strategy establishing that no shortcut exists to produce complex quantum entangled states. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) ChatGPT provided the key conceptual construction allowing researchers to disprove James Clerk Maxwell's historical rule on electrodynamic force equilibrium points. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) Claude derived a theoretical model for laser-exposed silicon membranes that closely matched physical experimental observations. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) AI models are handling complex physics engineering workflows, such as ion shuttling code for quantum computing and magnetic field selection for fusion stellarators. - [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) Physicists and mathematicians hold contrasting attitudes toward AI: mathematicians view it with existential dread, whereas physicists treat it as a practical calculation shortcut.

## Detailed walkthrough ### [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) AI's Rapid Takeover of Mathematics

Artificial intelligence has rapidly conquered mathematics, and the pace of development has become so overwhelming that it is hard to keep up. In recent months, an array of long-standing mathematical problems have fallen to cutting-edge models. Claude helped uncover a counterexample to the Jacobian conjecture, a problem that had remained stubbornly open since 1939. OpenAI announced ten separate results from its unreleased Astra model, with each result purportedly either solving or advancing a major open problem. Another group leveraged ChatGPT to disprove a 30-year-old conjecture, while Anthropic announced that Claude had achieved significant progress on the Riemann zeta function, inching closer toward the celebrated Riemann hypothesis.

Most mathematicians now concede that large language models are capable of doing serious, rigorous mathematics. With the automation of mathematics appearing increasingly imminent in the near future, early-career researchers and students find themselves caught in a volatile mix of depression and anger: - Graduate student Tasmin Chu declared that mathematics is facing an existential crisis and demanded an immediate moratorium on employing LLMs to prove novel theorems. - Assistant Professor Max Weinreich published a preprint arguing for total opposition to the integration of AI within the mathematical discipline. - Postdoc Alexis Marchand condemned LLMs as fundamentally harmful due to their massive contribution to the climate crisis. - PhD student Kerwin Hampshire reported feeling as though he were going insane, undergoing a profound spiritual crisis over these automated breakthroughs.

Sabine wryly observes that mathematicians have managed to "independently derived the same existential crisis by four different methods." Taking an equally grim view, Assistant Professor Daniel Litt gave a talk addressing the end of mathematics, warning that AI will ultimately flood academia with mathematically verified results that nobody will read, nobody will care about, and nobody will benefit from—reducing the discipline to an exaggerated parody of academic business as usual.

[02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) Physics Follows: Calculations, Proofs, and the arXiv Surge

If artificial intelligence is capable of guiding complex mathematical proofs, it naturally possesses the capacity to handle theoretical physics calculations. Sabine notes that she has tested this herself, and many theoretical physicists are quietly doing the exact same thing: feeding old research ideas that they never had time to explore into AI models and obtaining results solid enough for peer-reviewed publication. Peter Woit tracked submissions to the arXiv preprint server under the high-energy physics theory category (`hep-th`) and found that submission volumes spiked by roughly 30% in May and June compared to previous years.

```mermaid flowchart TD accTitle: AI Adoption Pipeline in Physics Research accDescr: Flowchart illustrating how physicists utilize AI models across theoretical, experimental, and engineering domains.

subgraph Inputs ["Research Concepts"] A["Unfinished research ideas and mathematical conjectures"] end

subgraph Models ["LLM Engines"] B["ChatGPT"] C["Claude"] end

subgraph Applications ["Physics Domains"] D["Quantum Entanglement (Adesso proof)"] E["Electrodynamics (Refuting Maxwell's rule)"] F["Experimental Optics (Silicon membrane modeling)"] G["Engineering & Fusion (Stellarator design & ion routing)"] end

A --> B A --> C B --> D B --> E C --> F B --> G C --> G ```

Concrete instances of AI-driven physics research have already emerged across diverse subfields: - **Quantum Entanglement:** Gerardo Adesso at the University of Nottingham sought to determine whether an easier path exists to generate complex entangled states. Using ChatGPT, which Adesso noted "made a central contribution to developing the proof strategy," he successfully proved that no such shortcut exists, subsequently verifying every algebraic step manually. - **Classical Electrodynamics:** Centuries ago, James Clerk Maxwell proposed a rule predicting the maximum number of equilibrium points where electric forces cancel out among randomly distributed charges (analogous to gravitational Lagrange points). Three researchers recently proved Maxwell's rule wrong, openly crediting ChatGPT with suggesting the core conceptual construction behind their counterexample. - **Experimental Condensed Matter/Optics:** Researchers investigating the behavioral response of microscopic silicon membranes exposed to laser light utilized Claude to derive an explanatory theoretical model, which demonstrated remarkable precision when matched against experimental observations. - **Applied & Computational Physics:** Additional publications have used LLMs to conduct symbolic and high-precision numerical computations in high-energy physics, run numerical algorithms in quantum theory, write specialized code for shuttling ions inside quantum computing architectures, and deploy ChatGPT agents to evaluate and select promising magnetic field configurations for fusion stellarators.

The cultural reaction among physicists stands in stark contrast to the existential dread gripping mathematicians. Rather than falling into philosophical despair, physicists have treated the models purely as labor-saving tools. As Sabine summarizes the divide: "Mathematicians see artificial intelligence and ask whether mathematics will survive. Physicists ask whether it can finish the calculation by Friday."

[05:25](https://www.youtube.com/watch?v=utu5YACZbPE&t=325s) Personal Data Privacy and Incogni

Sabine transitions to discussing personal online privacy, recalling an alarming incident a few years prior when a debt collection agency sent her threatening legal notices over an online order she had never placed. The ordeal, caused by identity theft, made her fiercely vigilant about guarding her personal data. Websites routinely harvest user identities, geolocations, and browsing histories to sell to commercial data brokers. While legal avenues exist to force data brokers to expunge private records, manually tracking down every company is extraordinarily time-consuming.

She introduces the video's sponsor, Incogni, a service designed to automate data removal requests across major data aggregators. In addition to handling recurring suppression notices and reporting progress back to the user, Incogni features a custom removal option that allows users to submit specific problem websites directly for automated opt-out processing.

## Notable quotes > [00:00](https://www.youtube.com/watch?v=utu5YACZbPE&t=0s) independently derived the same existential crisis by four different methods — Sabine Hossenfelder

> [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) made a central contribution to developing the proof strategy — Gerardo Adesso

> [02:17](https://www.youtube.com/watch?v=utu5YACZbPE&t=137s) Mathematicians see artificial intelligence and ask whether mathematics will survive. Physicists ask whether it can finish the calculation by Friday. — Sabine Hossenfelder

## Mentioned Alexis Marchand, Anthropic, Astra, arXiv, ChatGPT, Claude, Daniel Litt, Gerardo Adesso, Incogni, Jacobian conjecture, James Clerk Maxwell, Kerwin Hampshire, Max Weinreich, OpenAI, Peter Woit, Riemann hypothesis, Riemann zeta function, Sabine Hossenfelder, Tasmin Chu, University of Nottingham

## Video description Take back your personal data with Incogni! Use code Sabine at the link below and get 60% off annual plans: https://incogni.com/sabine

Artificial intelligence has eaten mathematics. In the past few months alone it turned up a counterexample to a conjecture open since 1939, disproved a thirty-year-old result, and made real progress on the Riemann hypothesis. Young mathematicians are somewhere between depression and anger about it. Physics is next, and almost nobody is talking about it. Let's take a look.

#physics #maths #ai #science #sciencenews #tech #technews

👕T-shirts, mugs, posters and more: ➜ https://sabines-store.dashery.com/ 💌 Support me on Donorbox ➜ https://donorbox.org/swtg 👉 Transcript with links to references on Patreon ➜ https://www.patreon.com/Sabine 📝 Transcripts and written news on Substack ➜ https://sciencewtg.substack.com/ 📩 Free weekly science newsletter ➜ https://sabinehossenfelder.com/newsletter/ 👂 Audio only podcast ➜ https://open.spotify.com/show/0MkNfXlKnMPEUMEeKQYmYC 🔗 Join this channel to get access to perks ➜ https://www.youtube.com/channel/UC1yNl2E66ZzKApQdRuTQ4tw/join 📚 Buy my book ➜ https://amzn.to/3HSAWJW

[Source](https://www.youtube.com/watch?v=utu5YACZbPE)

中文简报

arXiv理论论文提交量激增30%，物理学家借助AI开展研究在自动化定理证明让数学家陷入生存危机之际，理论物理学家正积极拥抱人工智能以加速经同行评审的计算。 Sabine Hossenfelder探讨了包括Claude和ChatGPT在内的现代大型语言模型如何迅速攻克重大数学里程碑，例如找出数十年前猜想的反例以及在黎曼zeta函数上取得进展。这些自动化成果在青年数学家中引发了普遍的生存恐慌、愤怒以及暂停AI证明的呼声。相比之下，物理学家则务实地采用AI来推导复杂证明、反驳经典电动力学规则、建模实验光学并优化聚变仿星器设计。Hossenfelder强调，尽管数学家担忧该学科的哲学未来，务实的物理学家却乐于将繁琐的计算自动化以加快论文发表。人工智能模型已取得重大数学突破，不仅发现了一个1939年猜想的反例，还在黎曼zeta函数的研究上取得了进展。 (0:00) 数学领域的快速自动化在青年数学家中引发了生存危机，并导致呼吁暂停AI生成证明的声音不断出现。 (0:00) 理论物理学界也正经历类似的转变，随着研究人员将积攒的想法输入LLM，arXiv高能物理理论的投稿量激增了约30%。 (2:17) 与争论学科生存问题的数学家不同，物理学家务实地采用AI工具来解决从量子纠缠到聚变仿星器设计等一系列问题。 (2:17)

Reading summary

Artificial intelligence has rapidly expanded from solving long-standing open problems in mathematics to executing complex calculations across theoretical and applied physics, according to Sabine Hossenfelder. While this automation has triggered existential dread and calls for moratoria among mathematicians, physicists have largely integrated large language models as practical labor-saving tools. AI Breakthroughs and Existential Crisis in Mathematics (0:00) Large language models have recently achieved breakthroughs on major mathematical challenges that resisted human solution for decades. Anthropic's Claude assisted in uncovering a counterexample to the 1939 Jacobian conjecture and made notable progress on the Riemann zeta function toward the Riemann hypothesis. Concurrently, OpenAI reported ten advancements or solutions to open problems using its unreleased Astra model, while another research team used ChatGPT to disprove a 30-year-old conjecture. These milestones have led most mathematicians to concede that AI can now perform serious, rigorous mathematical work. The prospect of automated mathematics has provoked severe distress among early-career researchers and students. Reactions include graduate student Tasmin Chu demanding an immediate moratorium on using LLMs to prove novel theorems, Assistant Professor Max Weinreich advocating total opposition to AI integration, postdoc Alexis Marchand pointing to severe climate costs, and PhD student Kerwin Hampshire experiencing personal spiritual crisis. Taking a broader view of academic incentives, Assistant Professor Daniel Litt warned that the technology threatens to flood academic journals with machine-verified proofs that no one will read, care about, or benefit from. Empirical Integration Across Theoretical and Applied Physics (2:17) Because theoretical physics relies heavily on mathematical calculation, researchers are increasingly feeding unresolved research concepts into LLMs to generate publishable findings. Tracking submissions to the arXiv preprint server's high-energy physics theory category (hep-th), Peter Woit identified an approximate 30% surge in submission volumes during May and June compared to previous years. Concrete implementations span theoretical formulations, experimental design, and engineering workflows. At the University of Nottingham, Gerardo Adesso credited ChatGPT with a central contribution to the proof strategy showing that no shortcut exists to create complex entangled quantum states, verifying each algebraic step by hand. In classical electrodynamics, researchers used ChatGPT to construct the counterexample that refuted James Clerk Maxwell's historic rule on electric force equilibrium points. In applied fields, researchers used Claude to derive predictive models for laser-illuminated silicon membranes, while others deployed LLMs for symbolic high-energy calculations, quantum ion-routing algorithms, and candidate selection for fusion stellarator magnetic configurations. Disciplinary Divide: Philosophical Anxiety Versus Pragmatic Adoption (2:17) The rapid encroachment of AI has revealed a stark cultural divide between academic disciplines. Where mathematicians express profound anxiety regarding the intellectual purpose and survival of their field, physicists have treated machine learning models primarily as calculation accelerators to be deployed without philosophical hesitation. Hossenfelder characterizes this contrast by noting that while mathematicians question whether their discipline will survive artificial intelligence, physicists focus strictly on whether the model can complete their calculations by Friday. The expanding capability of language models across pure mathematics and physics demonstrates that research automation is already underway, shifting the primary academic challenge from generating mathematical proofs to coping with the volume of machine-assisted output.

Source quotes

[0:00] Sabine Hossenfelder: independently derived the same existential crisis by four different methods [2:17] Gerardo Adesso: made a central contribution to developing the proof strategy [2:17] Sabine Hossenfelder: Mathematicians see artificial intelligence and ask whether mathematics will survive. Physicists ask whether it can finish the calculation by Friday.

Original transcript

[0:00] Artificial intelligence has eaten maths and now it's coming for physics. And the [0:05] interesting bit is no one's talking about it. So, let's talk about it. In maths, things have [0:11] been happening so fast I've given up trying to keep you updated. Since we last talked about it, [0:17] Claude Fable 5 helped find a counter example to the Jacobian conjecture, which I never heard of, [0:24] but can now tell you that it had been open since 1939. Then OpenAI announced 10 results from its [0:31] still unreleased Astra model, each of which either solved or advanced a long-standing open problem. [0:37] At least that's what they say. Another group used Chad GPT to disprove a 30-year-old conjecture that [0:43] no one had heard of previously either. And earlier this month, Anthropic announced that Claude had [0:49] made a major advance on the Riemann Zeta function. So working towards the Riemann hypothesis, [0:55] finally a topic that even I have heard of. I think at this point most mathematicians have [1:00] accepted that large language models are now capable of serious math. It seems likely to [1:06] me that mathematics will be entirely automated in the near future and especially young people just [1:12] entering the field are caught between depression and anger. Graduate student Tasmin Chu thinks that [1:19] mathematics is in an existential crisis and calls for a moratorium on using LLM to prove [1:26] novel mathematical theorems. Assistant professor Max Weinreich has written a preprint in which he [1:32] makes the case for total opposition to the use of artificial intelligence in mathematics. Postdoc [1:39] Alexis Marchand argues that LLMs are evil because they contribute to the climate crisis and PhD [1:45] student Kirwin Hampshire thinks he's going insane and is suffering a profound spiritual crisis [1:52] due to these developments. Mathematicians have independently derived the same existential crisis [1:58] by four different methods. Assistant Professor Daniel Litt, meanwhile, has given a more or less [2:05] serious talk about the end of mathematics in which he belabours that AI will lead to the production [2:11] of mathematically correct results that no one will read, no one will care about, and no one will [2:17] benefit from academic business as usual. But what about physics? If AI is now good enough to lead [2:23] maths proofs, then it should also be good enough to do physics calculations. In fact, it is. I know [2:29] this because I've tried. I also know that a lot of theoretical physicists are now doing the same [2:34] thing, taking research ideas that they haven't had time to follow through and put them into AI [2:41] to see what comes out. And sometimes the result is verifiably good enough for publication. Peter [2:47] Woit has looked at the number of submissions to the arXiv preprint server in the category hep-th, [2:54] that's high energy physics theory, and finds that it's up about 30% in May and June compared to [3:02] previous years. One of the recent examples comes from quantum physics. Physicists in this area have [3:08] long wondered if there's an easier way to create some complicated entangled states. Gerardo Adesso [3:15] at the University of Nottingham has now proved that the answer is no. And he did so with the [3:20] help of ChatGPT 5.6 about which he says the AI made a central contribution to developing the [3:27] proof strategy. He then checked every step of the final proof himself. Another recent example goes [3:33] all the way back to James Clerk Maxwell. Imagine a collection of electric charges in random places. [3:40] There will be some places where their electric forces exactly cancel kind of like Lagrange points [3:46] in gravity. Maxwell proposed a rule for how many such places you could get from a given number of [3:52] charges. Three researchers have now shown that Maxwell's rule is wrong and their paper says [3:57] the idea behind this construction was suggested again by Chad GPT 5.6. Those two examples are [4:04] still very mathematical. So here is one involving an experiment. These two researchers were studying [4:10] how tiny silicon membranes behave if you shine on them with a laser. For this they did measurements [4:16] and developed a model about which they write that it was derived with the help of Claude [4:22] Fable 5. The model has excellent agreement with the observations. I could go on like this for [4:28] a while. There's a paper in high energy physics that use several large language models to carry [4:33] out the symbolic and high precision numerical computations. Another paper in quantum physics [4:39] used claude for numerical calculations. Another team had Claude write software for moving [4:45] ions around inside a quantum computer and in fusion research Chad GPT agent was put in charge [4:51] of deciding which stellarator designs to try next. Physicists have clearly picked up on the LLM maths [5:00] revolution, but it's happening in a completely different way. They just organically integrate [5:06] the models as tools and move on. I also haven't yet seen the same desperation as I have seen [5:12] among young mathematicians. Mathematicians see artificial intelligence and ask whether [5:18] mathematics will survive. Physicists ask whether it can finish the calculation by Friday. A few [5:25] years ago, I received some ugly letters from a debt collection company. They threatened to sue [5:33] me for ordering something online and then not paying it. Turned out that someone had stolen [5:40] my private data. I managed to get out of this, but since then I've become very protective of my [5:46] personal information, which is why I've signed up to Incogn who've been sponsoring this video. Each [5:52] time you open a website, it'll try to collect data about who you are and where you are and what other [5:59] websites you've visited. Many firms then make money by selling your private information to data [6:05] brokers. Most countries have laws against that and you can ask for your data to be removed. But doing [6:11] this takes up a lot of time. Incogn automates the process of getting you out of those databases. [6:18] You sign up and they'll contact the big sinners request that your personal details be removed. [6:25] They'll keep on doing that and if you want send you updates about the progress they're making. [6:30] They now also have a custom removal feature where you can submit specific websites. I'll [6:38] be using this a lot. And Cognney has really made my life better. If that sounds like something you [6:44] could need to use my link, Zabina or the custom link in the info because the first [6:50] 100 people to use it will get 60% off. Stay safe. Thanks for watching. See you tomorrow.