Meu primeiro paper na @GoogleDeepMind
Colocamos dezenas de agentes do Gemini 4 Carbon para trabalhar nas equações diferenciais de dispositivos de fusão nuclear. Foram descobertas novas soluções exatas, incluindo uma que se sabia existir desde 1970, mas ninguém tinha encontrado.
wait, google deepmind researchers just used gemini 4 carbon agents to make a mathematical breakthrough in fusion research.
they've derived exact solutions for the magnetic fields that keep extremely hot plasma confined inside fusion reactors.
previously researchers could only find exact mathematical solutions for stellarator magnetic fields with a maximum of TWO repeating sections.
now they've derived THREE new families of exact solutions, extending these mathematical results to stellarators with ANY number of repeating sections around the reactor.
back in 1970, physicist dieter lortz proved that this particular type of magnetic equilibrium could exist.
now 56 years later we finally have an explicit formula for it.
these solutions can now serve as benchmarks for testing computer simulations of more complex stellarator magnetic fields, with 3 or more repeating sections.
the researchers guided the gemini 4 carbon agents through the research and independently checked the mathematical derivations.
this is still a preprint and not peer reviewed.
i genuinely don't think people realise how quickly ai is moving into seriously difficult theoretical physics problems.
The 3D codes that design fusion stellarators need exact solutions to check accuracy. Existing solutions covered at most 2 repeating segments, but real stellarators have 3 or more.
With Gemini 4 Carbon agents working together, we found solutions for any number N of segments.
Gemini 4 Carbon is now available.
It’s our strongest model, built to dig deeper, handle long-horizon tasks, and multi-step reasoning.
If Argon is great for complex tasks, Carbon is the model you bring in for your hardest problems.
we're so damn cooked 😭
google's gemini 4 argon just did something actually crazy for fusion energy
in 1967 a physicist named harold grad guessed that the magnetic fields inside twisty fusion reactors could never be solved exactly.
people believed him for 59 years
last month two papers finally proved him wrong. one of them used openai's GPT-6 Astra
then google deepmind gave that paper to a team of argon agents and they found a whole new family of answers nobody had.
the first answers showed up in a few hours.
the paper literally says "promising candidates appeared within hours"
NOT YEARS.
HOURS!!!!
the scientists checked every step by hand and it held up
59 years of humans vs a few hours of AI????
it's over. ai won.
Below is an animation (thanks Gemini!), comparing Matt’s stellarator-symmetric solution (left) with one of our new non-stellarator-symmetric equilibria (right).
It’s truly an impressive model, and I’m excited to see what else it can do in plasma physics! You can read more about Gemini Argon in the blog post: goo.gle/3W85sXj
I am still amazed by how Argon zero-shot solved this. Within hours after prompting, it found the mathematical generalizations, verified them symbolically, helped draft the manuscript and built interactive dashboards and animations to scan equilibrium parameters.
These solutions provide novel analytical benchmarks for testing numerical stellarator simulation codes in their full non-stellarator-symmetric regimes.
Using Google Antigravity’s multi-agent/teamwork-preview mode, we found two such families of solutions! They build directly on Matt’s work—generalizing both of his families (constant rotational transform and sheared rotational transform) and including them as special cases.
Matt’s counterexamples still had "stellarator symmetry," which is a discrete 180-degree rotation symmetry. At the end of his paper, he posed a follow-up question asking whether closed-form non-stellarator-symmetric counterexamples exist.
These are exact analytic solutions to the ideal 3D magnetohydrodynamic equilibrium equations (isomorphic to steady incompressible 3D Euler flow). In 1967, Harold Grad conjectured that smooth, non-axisymmetric solutions with non-constant pressure do not exist
We’ve found 2 new families of counterexamples to the long-standing (nearly 60-year-old) Grad’s conjecture in plasma physics using Gemini 4 Argon in our new paper, with Jonathan Citrin, Brendan Tracey, Matt Landreman at the University of Maryland: arxiv.org/abs/2610.07304!