GPT-6 Astra Pro, followed by teams of Gemini 4 agents, produced plasma equilibria that the physicist Harold Grad had deemed impossible, for less than $3 per problem.
In 1967, mathematician Harold Grad had argued that a plasma in equilibrium, with variable pressure and well-attached magnetic surfaces, could only exist smoothly with additional symmetry, as in a tokamak. The stellarators, whose twisted geometry did not have this symmetry, therefore seemed excluded. On September 22, Matt Landreman (University of Maryland) published two families of smooth three-dimensional solutions that contradict this conjecture: they were discovered with GPT-6 Astra Pro, and Landreman checked each equation by hand.
Agents in a few hours
A second publication, on October 5, co-signed by researchers from Google DeepMind and Landreman, entrusted the research to teams of Gemini 4 agents: promising candidates appeared in a few hours. These are the first balances of this type obtained in explicit form, which can serve as a reference for testing the design codes of stellarators. On the TheoPhysEval test bench, the average cost remains below $3 per problem, and $116.28 for all 50 API tariff issues.
This work does not say anything about the timing of a future reactor, and it is always humans who choose the questions to ask. Landreman believes that other exact solutions remain to be discovered.
Read the full article on the original source
Go to the source →