Key Moments

AI Is Breaking Math. Will It Solve Physics Next?

PBS Space TimePBS Space Time
Education5 min read25 min video
Nov 17, 2026|287,615 views|5,863|1,000
Save to Pod
TL;DR

AI is solving complex math and physics problems, but currently lacks the "scientific taste" and conceptual leaps needed to formulate truly novel research questions.

Key Insights

1

OpenAI's AI claimed a proof for the singularity aspect of the Navier-Stokes Millennium Prize problem in September 2026.

2

AI systems like Google's Gemini DeepMind, combined with tree search, have found new analytic approaches to problems involving cosmic strings, predicting testable gravitational radiation.

3

Representation learning in AI, exemplified by Carleo and Troyer's 2017 work on quantum states and FermiNet, allows neural networks to learn mathematical encodings of physical systems, reducing the need for human specification of variables.

4

AI Feynman, a hybrid system, successfully recovered 100 physics equations from numerical data by combining neural networks for structure exposure, dimensional analysis, and symbolic regression.

5

Systems like Symbolic Pregression and Pixel2Phy can start from raw video, learn hidden coordinates, and discover governing equations, effectively finding representations and laws simultaneously.

6

AI can design novel experimental apparatus, with systems like Melvin (2016) and Urania (2025) finding unconventional quantum optics and gravitational wave interferometer configurations that outperform human baselines.

AI's encroaching on mathematics and theoretical physics

The field of mathematics is undergoing an "AI-driven existential shift," prompting questions about whether physics will follow and when the first true AI physicist might emerge. OpenAI's September 2026 announcement of an AI system producing a proof for the singularity aspect of the Navier-Stokes existence and smoothness problem — one of the seven Millennium Prizes — highlights AI's growing capabilities. While the proof requires verification and may have been influenced by human work, it signifies a remarkable achievement. The Navier-Stokes equations, crucial for fluid mechanics, describe chaotic phenomena and hint at singularities, making them notoriously difficult to solve. Although the singularity itself is non-physical at atomic scales, AI's involvement in such complex mathematical proofs blurs the lines between pure mathematics and physics, raising anticipation and concern among mathematicians and physicists alike.

AI assisting in physics problems and law discovery

AI is increasingly contributing to physics research beyond pure mathematics. An AI-assisted system, combining Google's Gemini DeepMind with tree search, offered new analytical approaches to problems involving cosmic strings, hypothetical early-universe structures. This work predicted testable gravitational radiation, moving closer to physics than purely mathematical proofs. However, the AI's role was primarily in solving a complex integral, with human physicists still responsible for hypothesis generation, identifying measurable consequences, and interpreting results. This pattern highlights a recurring theme: AI excels at calculation and mathematical formalization, but human insight remains critical for the broader scientific process, including framing questions, designing experiments, and integrating findings into existing knowledge.

Representation learning and inductive bias in AI models

Early machine learning in physics primarily involved pattern recognition and calculation acceleration. However, the advent of representation learning marked a significant transition, allowing AI to learn mathematical encodings (representations) of problems itself, rather than relying on physicists to pre-specify variables and equations. In quantum mechanics, for instance, neural networks like those used by Carleo and Troyer in 2017 could represent complex many-body quantum states without explicit mathematical specification. FermiNet further advanced this by baking known physics, such as electron spin statistics, into a deep neural network while allowing flexibility to learn the rest of the wave function. This strategy of providing trusted physics principles (inductive bias) and letting AI learn the unknown parts is a recurring theme, though it raises questions about AI's ability to challenge fundamental assumptions when those biases are incorrect.

AI Feynman and symbolic regression push discovery boundaries

AI Feynman, a hybrid system, demonstrates AI's ability to uncover underlying mathematical forms from raw data. By combining neural networks for pattern identification, dimensional analysis for unit consistency, and crucially, symbolic regression to search for explicit mathematical expressions, it successfully recovered 100 physics equations from a benchmark dataset. Symbolic regression is key because it avoids shoehorning relationships into predefined structures. More advanced systems like Symbolic Pregression and Pixel2Phy go further, starting from raw video data, learning hidden coordinates where motion is simple, and then discovering governing equations. These methods favor simplicity (Occam's razor) and can find representations and laws simultaneously, sounding more like a physicist's work, though human invention and verification are still necessary.

AI in experimental design and execution

Beyond theoretical work, AI is also making inroads into experimental design. Systems like Melvin (2016) used combinatorial search to explore quantum optics experiment configurations, uncovering unusual setups missed by humans. More recently, Urania (2025) applied advanced methods to design gravitational wave interferometers, yielding unconventional configurations that outperformed human-optimized designs. While AI can propose and optimize experimental architectures, humans still drive the core scientific process: formulating hypotheses, interpreting outcomes, and integrating new knowledge. The ability of AI to perform many piecewise steps in a scientific project, from design to data analysis and even iterative refinement, is rapidly advancing, leading to the development of 'end-to-end AI scientists'.

The 'hard problem': generating truly novel questions

According to Battleday and Gershman's 'easy and hard problems' framework, AI excels at the 'easy' (though still challenging) task of performing assigned tasks and answering posed questions within a scientific process. However, the 'hard problem' — formulating truly novel questions that can lead to profound shifts in understanding, like Einstein's or Heisenberg's foundational inquiries — remains largely beyond AI's current grasp. While AI can generate 'combinatorial novelty' by reconfiguring existing knowledge, it struggles to break free from established scientific frameworks or 'inductive biases'. Recent studies, like Astro Insight (2026), show AI can suggest promising research directions, but these are often extensions of existing ideas rather than paradigm-shifting breakthroughs.

The role of scientific taste and understanding

Human physicists possess 'scientific taste' — an instinct for novel questions and promising research paths that integrates deep knowledge with curiosity, imagination, and skepticism. While AI can perform many individual components of this process, it currently lacks the ability to synthesize them holistically. Perhaps the most enduring human role in science is understanding the meaning of new results within the broader context of all known knowledge. Even if an AI were to discover a 'theory of everything,' its significance would be lost if humans could not comprehend its implications and integrate it into our worldview. Therefore, AI is unlikely to fully solve physics independently; its ultimate value lies in augmenting human understanding and directing scientific progress for human comprehension.

Common Questions

Yes, OpenAI announced in September 2026 that an AI system produced a proof for the Navier-Stokes existence and smoothness problem, one of the Millennium Prizes, though this claim is still under verification.

Topics

Mentioned in this video

More from PBS Space Time

View all 42 summaries

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