Juni
👤 SpeakerAppearances Over Time
Podcast Appearances
It shapes and stabilizes that plasma in real time using learned policies instead of hand-tuned control rules.
In October of 2025, Google DeepMind and the Commonwealth Fusion Systems, or CFS, announced a partnership to bring similar AI techniques into Spark, a next-generation fusion device that aims to pave the way to a commercial reactor called ARK.
Now, did they choose that because of Iron Man?
I don't know, but it sounds cool to me.
Maybe it was the other way around.
Now, that partnership just did an update in mid-November, which laid out how Spark is being built now, with AI expected to play a role in operations and optimizations.
So when you put it together, this is an early real-world example of AI not just analyzing data, but sitting inside the control loop of a critical physical system.
And it lines up with what we're seeing in grids, factories, and self-driving labs in other areas of science.
Now,
Here's a quick primer on the physics side of things.
Most of the serious private fusion efforts today use a tokamak, which is a donut-shaped device that confines a very hot, electrically charged gas called plasma using strong magnetic fields.
Think lightsaber.
The plasma needs to stay in a tight ring away from the walls.
If it touches the vessel or becomes unstable, the shot ends and you lose the chance to make useful energy or do good measurements.
Historically, engineers build controllers from physics equations and then carefully tune them for a few standard operating shapes.
It works, but it limits how aggressively you can push the machine and how much of the design space you can explore.
If this looks like a classic control problem in physics, it is.
This is where the AI actually comes in.
In the TCV work, researchers built a fast simulator of the plasma and let a deep reinforcement learning agent interact with this simulated tokamak.
It tried small changes to the magnetic coils, got rewarded when the plasma matched a desired shape and stayed stable, and slowly learned a policy for which actions to take at each moment.