keyhell
CPO @ Next Step Fusion
- Joined
- Nov 18, 2026
Comments
[Weekly] What are you building this week?
The timescale of tokamak plasma processes is 10^-5 s, so we have enormous data processing problems too, but so far I have mostly focused on validating our simulator (old scientific code, Fortran + Python) against the experimental data.
[Weekly] What are you building this week?
Both. We consider human time much more valuable, so if I can have an agent working instead of our physicist, it reduces the cost. Generating hypotheses is also great because AI can quickly study and retain so much data; experienced people beat it with experience and gut feeling, but still.
[Weekly] What are you building this week?
I'm teaching AI agents to run sophisticated simulations of nuclear fusion devices called tokamaks. Running simulations is not difficult, so an agent can figure it out by reading documentation and source code, but tokamaks are complex, and very often to achieve precise simulation a physicist or AI physicist needs to generate and test many hypotheses. AI can do this very well, but in many cases, it is not enough to have only one agent. Instead, I build chains of agents that generate hypotheses, review and prioritize them, independently estimate their impact, and so on.