2026 Conference on Physics and AI: Surya Ganguli

2026 Conference on Physics and AI: Surya Ganguli

by Surya Ganguli

Formal & Physical Sciences Physics PHPhysicsPHUMathematical
🎙 Surya Ganguli 👥 34K 📅 June 30, 2026 ⏱ 49 min 👁 195 📄 expert opinion 🧭 2026-08-03
Available in: English (current) Français

Keywords

AI for scienceinterpretabilityretina modelepilepsy digital twinscaling laws

Summary

Surya Ganguli, a professor at Stanford, delivers a keynote at the 2026 Conference on Physics and AI. He begins with a quote from Michael Atiyah, paraphrasing it to warn that AI might offer physicists powerful tools at the cost of conceptual understanding. He argues that while AI is valuable for practical applications, fundamental science must prioritize understanding. He proposes a framework where AI goes beyond prediction to provide explanations and guide experiments. He illustrates this with two case studies from his lab: modeling the retina and epilepsy. For the retina, they trained a CNN that accurately predicts responses to natural movies and generalizes to classic artificial stimuli, then used model reduction to derive interpretable models for each phenomenon, even discovering a new mechanism for the omitted stimulus response. For epilepsy, they built a digital twin of the epileptic brain using dynamical systems modeling, which accurately reproduces seizure dynamics. Analysis of the model revealed that seizures emerge from interactions between multiple brain regions, not a single locus. He also discusses the science of AI itself, presenting a theory of scaling laws for LLMs. He concludes by emphasizing the need for a science of AI to ensure we understand the tools we use.

201 words

Critical Evaluation

The talk is intellectually stimulating and presents a compelling vision for integrating AI into fundamental physics and neuroscience. Ganguli’s central thesis—that AI must move beyond prediction to provide conceptual understanding—is well-argued and timely. He supports this with concrete examples from his own research, demonstrating a rigorous approach to model interpretability. The retina work is particularly impressive: by training a deep network to predict responses to natural movies and then systematically simplifying it, they not only reproduced decades of experimental findings but also generated a new hypothesis. This exemplifies the ‘model reduction’ strategy he advocates. Similarly, the epilepsy digital twin is a powerful demonstration of how AI can uncover emergent mechanisms in complex biological systems. The finding that seizures arise from multi-region interactions challenges conventional wisdom and highlights the value of interpretable models. However, the talk is not without limitations. As a conference keynote, it provides an overview rather than deep technical detail. The scaling laws theory is mentioned only briefly, and the audience is left wanting more specifics. Additionally, while Ganguli warns against the ‘devil’s offer’ of AI, he does not fully address the practical challenges of implementing his vision, such as the computational cost of model reduction or the potential for AI to reinforce existing biases in scientific inquiry. The talk’s strength lies in its conceptual clarity and the quality of the research presented. The sources cited are primarily the speaker’s own papers and the conference itself, which is appropriate for a keynote but limits the breadth of external validation. Overall, this is a high-quality talk that offers valuable insights for scientists seeking to harness AI while preserving scientific understanding.

271 words

Title / Content Match

The title accurately reflects the content: a talk by Surya Ganguli at the 2026 Conference on Physics and AI.

Quality & Reliability

8/10

The speaker is a prominent researcher in theoretical neuroscience and AI, presenting at a prestigious conference. The talk is well-structured, references specific research projects, and includes cautionary remarks about the limitations of AI. However, as a conference talk, it lacks detailed methodological descriptions and peer-reviewed citations for all claims.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Concurring Sources

Contribution & Novelties

The talk presents a clear framework for integrating AI into fundamental science, emphasizing the importance of interpretability and conceptual understanding. It showcases two concrete examples from the speaker’s lab: a retina model that generalizes to classic stimuli and yields new mechanistic insights, and an epilepsy digital twin that reveals emergent seizure dynamics. The talk also hints at a novel theory of scaling laws for LLMs, which could have broad implications.

Pour aller plus loin :

  • Model reduction in neuroscience — Relevant to the interpretability approach discussed.
  • Digital twin — Concept applied to epilepsy modeling.
  • Scaling laws for neural language models — Key paper on empirical scaling laws, related to the theory mentioned.

112 words

Radar Profile

The radar profile shows high scores in quality of information and reliability, reflecting the speaker's expertise and the conference setting. The quantity of information is moderate, as the talk covers several topics but not in exhaustive depth. The technical level is high, suitable for a specialized audience.

Reliability 8/10