Key Moments
🔬Top Black Holes Physicist: GPT5 can do Vibe Physics, here's what I found
Latent Space PodcastKey Moments
AI can now perform complex theoretical physics calculations, reproducing research-level papers in minutes and solving year-old puzzles, raising questions about future scientific training and discovery.
Key Insights
GPT-5 reproduced one of the speaker's best research papers, which took a long time to develop, in just 30 minutes.
The paper 'Single Minus Gluon Tree Amplitudes Are Non-Zero' was solved with AI assistance, a problem that had puzzled physicists for over a year.
A subsequent paper extended this result to gravitons, with GPT Pro performing the calculation, a task that would have taken significant human effort.
Co-host RJ noted that the AI's ability to identify mathematical tools like the directed matrix tree theorem for the graviton paper was a surprising and valuable insight.
The speaker believes that current AI models can produce papers as good as human-written ones, potentially leading to an archive inundated with submissions without careful steering and verification.
AI significantly reduces the time physicists spend confused by calculations and helps chart research paths faster by acting as a 'scout' for promising approaches.
AI's rapid ascent in scientific reasoning capabilities
The conversation with Alex Lupsasca, a renowned physicist, highlights a pivotal moment where AI has surpassed human capabilities in specific scientific tasks. Lupsasca recounts his personal journey from skepticism to full embrace of AI in theoretical physics. Initially, he believed AI was useful for mundane tasks like email but not for complex research. However, the release of ChatGPT 0.3 marked a turning point, exhibiting strong reasoning and mathematical abilities. The subsequent release of GPT-5 was a watershed moment, as it reproduced one of Lupsasca's most significant research papers—a feat that took him a considerable amount of time—in a mere 30 minutes. This experience led him to declare it the 'most important discovery of my lifetime,' compelling him to join OpenAI to explore these advancements further. He notes that many senior colleagues in the physics community are now also embracing these AI capabilities.
From puzzling problems to AI-driven solutions
A key example cited is a problem in theoretical physics concerning 'single minus gluon tree amplitudes' that had stumped experts for over a year. This problem, detailed in a paper titled 'Single Minus Gluon Tree Amplitudes Are Non-Zero,' involved understanding interactions between gluons, the particles that carry the strong nuclear force. It was previously thought that amplitudes with a single particle of opposite helicity (a 'single minus' configuration) were always zero due to fundamental principles. However, researchers discovered a loophole in generic conditions, suggesting these amplitudes might not be zero in a specific 'collinear' alignment of particles. While humans identified the loophole and the potential for non-zero amplitudes, deriving the actual simple formula proved exceptionally difficult, with manual calculations leading to complex, factorial-growing terms. AI, specifically ChatGPT, was instrumental in simplifying these calculations, identifying a special region in the 'phase space' where the amplitudes significantly simplified. The AI not only found a compact formula for the five- and six-point amplitudes but also conjectured a general formula for any number of gluons, a feat that would have taken immense human effort.
Extending AI's reach to quantum gravity
The success with gluons paved the way for applying similar AI-driven approaches to gravity, mediated by gravitons. In a paper titled 'Single Minus Graviton Tree Amplitudes Are Non-Zero,' the researchers aimed to extend the findings from the gluon case to gravitons. While gravity is mathematically more complex due to gravitons being spin-two particles, the AI, using the gluon paper as a 'seed' and the public GPT Pro model, successfully performed the graviton calculations. This involved understanding manipulations in appendices of the gluon paper and applying them to the gravity context. The AI identified crucial mathematical steps similar to human research, including the application of the directed matrix tree theorem, and independently derived the equations. What was particularly remarkable was that the AI, given the gluon paper and specific tweaks for gravity, was able to generate a draft of the paper that was very close to the final published version, with human effort primarily focused on verification and adding broader context, such as symmetry transformations.
The 'vibe physics' of AI-assisted discovery
Lupsasca describes the process of using AI for complex calculations as 'vibe physics.' This refers to the AI performing extensive calculations, exploring equations, and verifying intermediate steps much like a human physicist, but at an unprecedented speed and scale. He highlights how the AI could independently discover and apply relevant mathematical theorems, such as the directed matrix tree theorem for the graviton paper, which even surprised the human researchers. The human role shifted from performing arduous calculations to guiding the AI, asking the right questions, and meticulously verifying the AI's outputs. This collaborative process, where the AI acts as a highly capable, tireless assistant, dramatically accelerates the pace of scientific discovery.
Implications for scientific training and the future of research
The unprecedented capabilities of AI in theoretical physics raise profound questions about the future of scientific training. Lupsasca and his co-hosts discuss how traditional graduate training, which often involves arduous calculations as a rite of passage, may need to evolve. AI can now potentially solve problems that would have taken students months or even years. This shift means future physicists might learn to collaborate with AI, focusing on formulating the right questions, understanding concepts, and verifying results, rather than just performing calculations. The role of professors may also change, from assigning 'easy' problems to guiding students in leveraging AI tools effectively. While AI can eliminate much of the confusion and wasted effort in research, the core skill of knowing what questions to ask and identifying fruitful research directions remains a critical human contribution.
The bottleneck of writing and verification
A significant bottleneck identified in modern research is the process of writing up findings and verifying complex calculations. Lupsasca suggests that traditional paper formats might become obsolete, contemplating a future with interactive, AI-integrated scientific documents. The AI's speed in generating results means the human effort is now concentrated on verification. He notes that while current models are incredibly capable, ensuring their accuracy is paramount, especially as calculations become more complex. This necessitates advancements in formal verification methods and models that can more explicitly indicate their confidence levels in their outputs. The speed at which AI can now produce results also presents a challenge regarding 'AI slop'—unverified or incorrect submissions to archives—which the academic community needs to address by raising the bar for what constitutes a high-quality research paper.
Pushing the frontier of knowledge with AI
The current trajectory suggests AI will continue to rapidly advance, enabling breakthroughs in areas like quantum gravity. Lupsasca is optimistic that AI will provide 'superpowers' to human physicists, allowing them to tackle increasingly difficult problems. While AI can now solve problems that stumped experts for years, the ultimate challenge remains for AI to independently identify questions that have stumped entire communities for decades. This requires a level of creativity and intuition that is still evolving. However, by scaling up AI capabilities and refining how we prompt and steer these models, scientists are pushing towards a future where AI not only solves problems but also aids in discovering new conceptual leaps, much like the breakthroughs in the single minus amplitude papers.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Concepts
●People Referenced
Common Questions
Initially, AI was only useful for simple tasks like email over a year ago. However, with the release of ChatGPT 3, it became capable of complex math, and GPT-5 reproduced highly difficult research papers in minutes. More recently, AI has solved open problems that puzzled physicists for over a year.
Topics
Mentioned in this video
The university where Alex Lubyansky is a professor.
The European Organization for Nuclear Research, home to the LHC in Geneva.
An academic institution where Alfredo Guevara is a postdoc.
The university where David Skinner is a professor.
The university where Andrew Strominger is a professor.
A magazine that published a great article understanding and covering the AI physics discovery.
The university where an OpenAI event with IPAM (Institute of Mathematics) took place, where Alex Lubyansky talked to Terry Tao.
An institute of mathematics at UCLA that hosted an OpenAI event.
A language for formal verification mentioned in the context of mathematical proof.
The first really strong reasoning model mentioned by Alex Lubyansky, useful for mathematical research and capable of saving a lot of time.
An advanced AI model that was able to reproduce one of Alex Lubyansky's papers in 30 minutes, leading him to believe it would change research fundamentally.
An AI model that could simulate the SYK model in 10 minutes, a very technical quantum mechanics problem that research groups struggled with.
An earlier version of ChatGPT that conjectured the general formula for single-minus amplitudes, but couldn't quite prove it.
A company where co-host Brandon works, developing RNA therapeutics using AI.
The company where co-host RJ Honicky is CTO and founder.
The AI research company where Alex Lubyansky is a fellow and conducts research using GPT models.
A social media platform where the initial reception to GPT-5 was lukewarm, and later used by Alex Lubyansky to clarify the gluon paper findings.
A company associated with formal verification, whose co-founder (Korean Hong) would agree with the importance of automated verification.
A professor at Vanderbilt University and fellow at OpenAI, winner of the 2024 New Horizons Breakthrough Prize and IUPAP award, doing research at OpenAI pushing theoretical physics using GPT models.
Mentioned by the co-host as someone who had a similar realization about AI's capabilities as Alex Lubyansky.
His theory of relativity is one of the two basic principles of nature that physicists believe every law should respect.
A postdoc at the Institute for Advanced Study, co-author on the gluon paper, who did a lot of hard manual calculations that AI later simplified.
A professor at Cambridge University and co-author on the gluon paper.
A professor at Harvard, Alex Lubyansky's former advisor, and co-author on the gluon paper, who had been thinking about the problem for a year.
A particle physicist and co-author on the gluon paper.
One of the pioneers of Quantum Field Theory, who developed Feynman diagrams as a visual way to understand particle interactions.
A renowned mathematician who, in his view, has not yet been impressed by creative moves in math proofs from AI, seeing them as recombinations of obscure facts.
A British mathematician who studied tides, after whom 'Love numbers' are named.
A physicist whose work sometimes suggests that single-minus non-zero amplitudes are exceptions to a zero measure set.
A technical model in quantum mechanics and gravity that Codex was able to simulate effortlessly.
A general framework that describes physical forces of nature, accommodating both relativity and quantum mechanics, which is currently the best theory.
The particle of light, mediating electromagnetism, and described as having polarization.
Particles that mediate the strong nuclear force, binding the nucleus together; focus of a breakthrough paper where AI found their amplitudes are non-zero.
A simple formula developed in the 1980s by physicists Parke and Taylor for double-minus amplitudes, which the AI-discovered formula for single-minus amplitudes is an analog of.
Maximally helicity violating amplitudes, which the Parke-Taylor formula describes, now considered a bit of a misnomer due to the new findings.
Visual cartoons representing possible particle interactions, organized by complexity, that allow physicists to understand and compute quantum amplitudes.
A crucial mathematical application that the AI surprisingly used in solving the graviton problem, unknown to the human physicists involved.
Its irreducible representations define particles in quantum field theory.
A technical term in physics referring to coefficients that encode the strength of tidal response. Black holes famously have zero Love numbers, a fact linked to symmetry principles.
A specific spacetime metric for which GPT-5 was asked to find symmetries, forming part of a complex black hole problem.
Hypothetical particles that mediate gravity, which are the quantum of gravity; the subject of a follow-up paper where AI extended the gluon amplitude results.
More from Latent Space
View all 211 summaries
61 min⚡️ Competing with ChatGPT and Sierra, building a $10M ARR company — Yasser Elsaid, Founder, Chatbase
75 minCI/CD Breaks at AI Speed: Tangle, Graphite Stacks, Pro-Model PR Review — Mikhail Parakhin, Shopify
86 min🔬 Training Transformers to solve 95% failure rate of Cancer Trials — Ron Alfa & Daniel Bear, Noetik
49 min⚡️ How to turn Documents into Knowledge: Graphs in Modern AI — Emil Eifrem, CEO Neo4J
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free