AI just cracked a 60-year-old physics conjecture

Nuclear fusion relies on powerful magnetic fields to bottle up superheated plasma, and stellarators—with their wildly complex, twisted 3D fields—are among the most promising paths forward. But plasma physicists have hit a mathematical wall since 1967: can a smooth, stable plasma structure exist in 3D magnetic fields if pressure varies, without satisfying strict symmetry conditions? For nearly six decades, the consensus was no.

Two new papers shattered that long-standing belief. Researchers used GPT 6 Astra to unearth two brand-new counterexamples, proving that stable and smooth plasma configurations can thrive without the traditional symmetry requirements.

Breaking a 60-year theoretical dogma

Plasma is an ultra-hot gas of charged particles, and keeping it bottled up requires incredibly precise magnetic confinement. For decades, theorists assumed that without special symmetry properties, any variation in plasma pressure would destroy stability.

Instead of wrestling through the calculations by hand, researchers turned to GPT 6 Astra to explore the solution space. The AI bypassed decades of human intuition and surfaced two valid configurations that defy the old rule. It marks another massive win for AI in theoretical physics, coming right after OpenAI used machine models to prove fluid vortices can form singularities under Navier-Stokes conditions.

Why it matters

We think this highlights the sweet spot for frontier AI in pure science: disproving dogmas by surfacing counterexamples.

Human scientists frequently get trapped by intuitive assumptions, making counterexamples notoriously hard to stumble upon. But once an AI spots one, verification is relatively straightforward for humans. The division of labor works naturally.

For the fusion sector, ditching strict symmetry constraints could fundamentally liberate stellarator magnet designs. Engineers might no longer need to hit impossible geometric tolerances to build working reactors. AI is rapidly graduating from a code-generating assistant to an actual pioneer pushing the boundaries of physical theory.