On October 5, a paper titled "Analytic non-axisymmetric 3D MHD equilibria without stellarator symmetry" appeared with Matt Landreman of the University of Maryland among its authors. Google DeepMind promoted the work in a thread on X as proof that Gemini 4 Argon had helped break Harold Grad's 1967 conjecture. One awkward detail: by October 5, Grad's conjecture had already been broken, in two papers published a day apart, and the model in the room for the decisive result was not Gemini.
Grad, a physicist and mathematician at NYU, proposed in 1967 that you cannot have a smooth three-dimensional plasma equilibrium without some underlying symmetry in the magnetic field. The claim stood as a quiet obstacle for stellarators, the twisted doughnut reactors that confine plasma using deliberately asymmetric coils rather than simpler symmetric tokamaks like ITER. If Grad was right, there was a mathematical ceiling on how cleanly anyone could describe the physics those machines depend on. It was always a conjecture, never a theorem.
The record of who got there first is unusually well documented, because Landreman published his prompts. At 6:44 a.m. on September 10, he asked GPT-6 Astra Pro for an asymmetric magnetic cage with nested surfaces, zero divergence, helical field lines, and MHD force balance. The model returned a complete family of exact solutions in 20 minutes and 34 seconds. It even flagged its own weakness: the rotational transform was an integer, meaning field lines close on themselves after one loop.
At 4:05 p.m. the next day, Landreman sent the solution back and asked for the harder version, an irrational rotational transform so field lines never close. Astra spent 33 minutes and 37 seconds, opened by saying it could not do it, then proved why patching the first family was impossible, switched construction, and delivered a second family. It checked its own work by substituting the equations at 1,202 points with 64-bit automatic differentiation, then re-verifying with a fourth-order finite difference scheme.
Then the humans took over for roughly ten days. Landreman ran both families through the DESC stellarator equilibrium solver, re-derived the equations term by term in SymPy and Mathematica, and uploaded the paper to arXiv on September 22. The acknowledgments read: "These solutions were discovered using the artificial intelligence model GPT-6 Astra Pro, and part of the paper was also drafted by it. All equations have been manually confirmed by the authors." He also published the conversation as a PDF alongside the verification scripts. A separate paper by Javier Gomez-Serrano, Lukas Liehr and Mitchell Taylor landed within a day, and together they produced three families of counterexamples. Grad's ceiling was gone.
What Google's Paper Actually Adds
The Gemini paper is real work, and it does something the earlier results did not. The new solutions drop a residual stellarator symmetry constraint that survived in the earlier families, then introduce new parameters that generalize it. Set them to zero and you recover Landreman's original solutions, the signature of an extension rather than a first strike. That is a legitimate mathematical contribution. It is also a follow-up.
The method matters more than the marketing. According to chapter five of the paper, "Use of AI tools," the team ran Gemini 4 Argon as a small group of agents inside Google's Antigravity environment, feeding it Landreman's earlier papers and a list of unsolved problems. The agents proposed the structure; researchers critiqued, independently re-derived, and confirmed it held. Gemini also handled symbolic computation, wrote numerical verification code, and drafted parts of the manuscript. Google gets credit for disclosing this at all. Plenty of labs quietly leave the model out of the methods section.
Where the Headline Overreaches
The gap between what happened and how it was sold comes down to three things:
- Sequencing. The conjecture fell in September. Google's framing of a sixty-year problem falling in a launch thread erases the two weeks in which the result already existed, and the fact that the underlying solutions came from a competitor's model.
- Evidence standard. Landreman published prompts, transcripts, and scripts. Google has not published a blow-by-blow of how Argon reached its counterexamples. The claim rests on a social media thread and a methods paragraph.
- Milestone inflation. A conjecture being disproven is not a reactor. Confinement physics is one layer of a stack that also includes neutron-resistant materials, tritium breeding, and plant economics.
None of that makes the result worthless. Exact analytical solutions give designers something new: closed-form answers to check simulation codes against. Teams can run the workhorse VMEC solver or DESC on a problem where the right answer is already known, and see whether the code arrives there. Companies like Commonwealth Fusion Systems and Type One Energy spend heavily on designs validated by those codes. Fixed points for verification are genuinely useful, and far less exciting than a headline about sixty years of physics falling.
The Pattern, Not the Physics
Watch how this gets packaged. Gemini 4 Argon was first handed to vetted cybersecurity defenders through Google's Fairwind Program rather than paying customers, and roughly 10,000 scientists got free research access, a move that will reliably generate more stories like this one. The model leads most disclosed benchmarks and trails Claude Opus 5.5 on coding-agent tasks. A conjecture falling in a launch thread is a better advertisement than any benchmark chart, which is precisely why it gets framed that way.
The honest version is less cinematic and more interesting. A model generated candidate solutions in under an hour that a specialist had spent years circling. A human spent ten days verifying them. A second model generalized the result, and humans checked that too. That division of labor, models proposing and people proving, is the actual news. Grad's ceiling is gone, three families of counterexamples exist, and stellarator codes have new fixed points to test against. But the model that got there first was GPT-6 Astra, and no reactor came online because of any of it.
Comments