Magnetic control of tokamak plasmas through deep reinforcement learning

Feels like a major win. Would not have predicted seeing "droplet" shaped zones of plasma stability. Lyman Spitzer himself would be amazed at this. One of those AI results that begins to feel like the emergence of an alien intellect. And of course, possessing a general feedback control loop agent that could be applied to any domain begins to feel like DeepMind could point the thing at just about any domain ;)

This seems like one of those ‘the future is now’ moments.<p>Kid in museum in 2100: “look mom! It’s the first AI that wrestled a star on earth and won!”

Awesome to hear that RL for plasma confinement could work for fusion progress.