Key points
- A 59-year-old open question: In 1967, Harold Grad posed a classic conjecture that steady states of asymmetric toroidal plasmas satisfying magnetohydrodynamic (MHD) force balance do not exist. It bears directly on the theoretical foundations of fusion devices such as stellarators.
- 33 minutes: University of Maryland physicist Matt Landreman used GPT-6 Astra Pro to derive, in just 33 minutes, exact analytic solutions with asymmetric, three-dimensional nested magnetic surfaces, directly disproving the Grad conjecture.
- Fully public: The complete work, along with all prompts, has been posted on arXiv, showing a large language model's capacity for direct discovery in demanding theoretical-physics modeling, rather than after-the-fact explanation.
Analysis
This is not "AI-assisted computation" but AI-led theoretical derivation, from physical modeling to analytic solution in one pass. For the fusion community, proof that asymmetric steady states exist gives stellarator configuration optimization a new theoretical footing. For the AI community, it shows that frontier models can now handle symbolic derivations that call for deep domain intuition. 33 minutes versus 59 years: the contrast is itself the news.
Source: OpenAI Delivers on Day 3 Challenge: Full Rollout of GPT-6 (AI Daily 2026-10-09). This text is a summary for reference only.
