News · 2026-10-18

AI-assisted plasma counterexamples come with a priority caveat

Physicist Matt Landreman has published explicit mathematical counterexamples to Harold Grad’s 1967 conjecture about certain three-dimensional plasma equilibria, crediting GPT-6 Astra Pro with discovering the solutions. Landreman says he manually checked every equation. The result supplies exact examples useful for theory and numerical testing, while neither proving better reactor confinement nor establishing that AI produced the first counterexample.

Key facts

• Landreman presents two parameterized families of exact solutions, expressed with elementary functions.
• His preprint was submitted September 22, one day after a separate counterexample paper with Lean 4 certification.
• A follow-on submitted October 5 generalizes the families to break an additional symmetry.
• The primary source is Landreman’s full paper, with the earlier counterexample essential to the chronology.

What the old conjecture concerned

Magnetic-confinement research studies how magnetic fields can hold a hot plasma in suitable arrangements. One mathematical question is whether the governing equilibrium equations allow smooth, non-axisymmetric configurations with nonconstant pressure on a toroidal domain: broadly, a doughnut-shaped region without ordinary rotational symmetry.

Grad’s conjecture concerned an obstruction to such arrangements without special symmetry. This is a question about the existence and structure of mathematical solutions. It is related to the design of stellarators, but it is not equivalent to asking whether a reactor will produce useful power or hold a real plasma stably under operating conditions.

An analogy is finding an exact design that satisfies a set of bridge equations. That can overturn a claim that no such mathematical design exists. It does not establish that a bridge built from particular materials will withstand wind, fatigue, and every other condition outside those equations.

Explicit solutions make the result inspectable

Landreman’s paper is titled “Analytic toroidal 3D MHD equilibria and steady Euler flows with invariant surfaces.” The title captures both the plasma setting and a mathematical connection: the equations also correspond to steady incompressible fluid flow.

The two solution families specify the magnetic field, pressure, and flux surfaces explicitly in Cartesian coordinates. They are smooth over the toroidal domain. One family has a fixed rotational transform of two, a measure describing how field lines wind; the other has a profile that changes across the configuration.

The existence of explicit formulas matters. Researchers can substitute them into the equations and use them as known answers when testing numerical software. A simulation that fails to recover a known solution has revealed something about its own approximation or implementation. That value is concrete even before any claim about reactor performance.

The AI credit is clear; the discovery process is less visible

In the acknowledgements, Landreman credits GPT-6 Astra Pro with discovering the solutions and helping draft parts of the manuscript. He reports manually confirming every equation. Those are direct disclosures by the author, not an inference from a social-media post.

The paper does not provide the complete prompts, search history, number of failed attempts, or an independent reproduction of the model’s discovery process. Readers can therefore distinguish the mathematical object from the account of how it was found. Exact formulas can be checked even when the exploratory interaction that produced them is not fully documented.

This is also different from formal proof checking, in which a proof is expressed in a language a small trusted checker can verify. A manual equation check and a machine-certified proof can both be valuable, but they describe different verification processes.

The priority claim needs a correction

Javier Gómez-Serrano, Lukas Liehr, and Mitchell A. Taylor submitted Counterexamples to Grad’s conjecture on September 21. Their abstract describes smooth solutions on embedded solid tori with nested pressure levels and provides Lean 4 certification. Landreman’s paper followed on September 22 with different explicit analytic families.

The correct story is therefore not that an AI system was the first to disprove the conjecture. It is that Landreman reports AI-assisted discovery of particularly explicit counterexamples, within a sequence of mathematical developments that includes an earlier result.

On October 5, Opal Issan, Jonathan Citrin, Landreman, and Brendan Tracey submitted a generalization that adds a parameter breaking stellarator symmetry. They disclose using Gemini 4 agents to propose generalizations, followed by independent re-derivation and symbolic checks. Because Landreman is an author, this is follow-on work, rather than wholly independent validation of his original paper.

The result is useful without a fusion breakthrough headline

These are preprints. The dossier establishes author-reported checking and a technical extension, not completed peer review or experimental confinement gains. Public circulation in October also does not make the September submission a new October 18 release.

The significant pattern is a division of labor: a model proposes mathematical structure, while researchers check it and place it in the literature. The strongest caveat is not that the formulas lack value; it is that discovery credit, publication priority, mathematical verification, and practical reactor usefulness are separate claims that require separate evidence.

Primary source, verified: read the paper → (arXiv 2609.26742)

Ground Truth. (2026, October 18). AI-assisted plasma counterexamples come with a priority caveat. Ground Truth. https://groundtruth.day/news/landreman-ai-assisted-grad-counterexamples-priority-context.html

@misc{groundtruth:landreman-ai-assisted-grad-counterexamples-priority-context,
  title  = {AI-assisted plasma counterexamples come with a priority caveat},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/landreman-ai-assisted-grad-counterexamples-priority-context.html}
}

Topics: science · mathematics · ai-for-science · plasma-physics