Google's Gemini AI helped disprove a 60 year old fusion physics conjecture

Harold Grad raised this obstacle for stellarator physics in 1967, and it's stood unresolved ever since. Now two independent research efforts have broken it open, including one from a Google DeepMind team using Gemini 4 Argon, right as investors keep pouring billions into fusion.

Harold Grad, a physicist and mathematician at NYU, conjectured in 1967 that you basically can't have a smooth, three dimensional plasma equilibrium without some kind of symmetry in the magnetic field. For decades that stood as a quiet but real obstacle for stellarators, the twisted, donut shaped reactors that confine plasma using deliberately asymmetric magnetic coils, as opposed to the simpler, symmetric tokamak design used by ITER. If Grad was right, there was a mathematical ceiling on how cleanly you could describe the physics underpinning that confinement. Matt Landreman, a research professor at the University of Maryland who has spent years on stellarator optimization, found two new counterexamples that break that ceiling, working with OpenAI's GPT-6 Astra. Separately, a Google DeepMind team says Gemini 4 Argon helped it uncover further counterexamples of its own.

Google DeepMind laid out its own findings in a thread on X, framing them as one of several breakthroughs tied to Argon's launch, alongside work in quantum error correction and protein design. The broader fusion result isn't just a tidy math trick. Researchers now have three distinct families of counterexamples to Grad's conjecture, split across two independent arXiv papers that landed within a day of each other: one from Javier Gomez-Serrano, Lukas Liehr and Mitchell Taylor, and another from Landreman. These exact analytical solutions for stellarator magnetic field equations give physicists something they didn't have before: reliable, closed-form benchmarks to check whether the simulation codes used to design next-generation stellarators are actually getting the physics right.

That distinction matters more than it sounds. Stellarator design today leans heavily on numerical codes that approximate plasma equilibrium, because exact solutions were assumed, per Grad, to not really exist outside of narrow symmetric cases. Without an exact answer to check against, engineers are partly flying on faith that their simulations converge on something physically real. A verified analytical solution lets you run the simulation code on a problem where you already know the right answer, and see if it gets there. Commonwealth Fusion Systems depends on exactly this kind of simulation fidelity to justify reactor designs before committing hundreds of millions of dollars to hardware. So does Type One Energy, which is building a stellarator-based demonstration reactor in Tennessee.

Google hasn't published a blow-by-blow of Argon's reasoning process on its own counterexamples, but the company says the approach follows a pattern DeepMind has leaned on elsewhere: pair the model with domain experts and let it generate and check candidate solutions against the magnetohydrostatic equations until it turns up families that genuinely violate Grad's assumptions. Landreman's counterexamples, by contrast, came out of work with GPT-6 Astra, with Landreman manually verifying every equation. Researchers say solutions like these could be folded directly into verification suites for design codes like VMEC, the workhorse equilibrium solver much of the stellarator community already uses.

Google Restricts Its Most Powerful AI Model to Vetted Cyber Defenders First

Google launched Gemini 4 Argon on September 30 and gave unrestricted access first to vetted cybersecurity defenders through its Fairwind Program, not to paying customers. The model leads rivals on most disclosed benchmarks but trails Claude Opus 5.5 on coding-agent tasks, and Google says it's gating the release over the model's offensive cyber... - Google restricts Gemini 4 Argon to cyber defenders - how to access Google's Gemini 4 Argon model

Google says it's giving roughly 10,000 scientists free research access to Argon, according to reporting from EdTech Innovation Hub, a move clearly meant to generate more stories exactly like this one. Whether that's generosity or a very deliberate recruiting pitch for AI-assisted science probably depends on who you ask. Either way, a 60-year-old conjecture falling in a launch thread is a better advertisement than any benchmark chart.

Why this is also a Gemini versus Claude story

It's worth being blunt about something SF readers already track closely: this is Google's model, not Anthropic's, scoring the win. Anthropic has had its own run of AI-for-math moments this year. Donald Knuth reportedly had Claude Opus 4.6 crack a graph theory problem on Hamiltonian cycles in about an hour, after he'd spent weeks stuck on it himself. He called the solution "a dramatic advance in automatic deduction and creative problem solving." Separately, Anthropic says its Claude Mythos model independently solved the 1946 Erdős unit-distance conjecture in discrete geometry, shortly after OpenAI claimed the same result. Fields medalist Timothy Gowers called that one a milestone for AI-assisted mathematics.

Put those side by side and a real competitive axis appears. Anthropic's wins have clustered in pure mathematics: graph theory, discrete geometry, cryptography. Google's result is different. Grad's conjecture is the first of this current wave to land squarely inside an applied physics problem with a direct industrial customer, the fusion companies racing to get a commercial reactor running before the end of the decade. Frankly, that's the more commercially legible flex. A solved graph theory problem is a trophy for mathematicians. A verified stellarator benchmark is something a fusion startup's engineering team can plug into its simulation pipeline next quarter.

None of this means Grad's conjecture falling makes commercial fusion arrive any sooner. Confinement physics is one layer of a stack that also includes materials that survive neutron bombardment, tritium breeding, and the plain economics of running a power plant. But stellarator developers have spent decades designing partly in the dark on whether their simulation codes track reality. Now they have a few more fixed points to check against. They got them from a chatbot, with an actual plasma physicist checking its work.

Stripe's $7.5 Billion OpenRouter Deal Puts a Payments Company in Charge of AI Model Choice • Banks are locking AI agents out of accounts the same week labs admit they go rogue • Websites are locking AI crawlers out as agent bot traffic surges in 2026

This article is posted in AI News, check it out for more related stories.