Important note - lasers are relevant in ion traps or neutral atom quantum computers. The biggest players are using superconducting as a base - which doesn't need lasers, but instead near absolute zero temperatures. Obviously that's insanely expensive, but also easier in some ways; so alternatives still have some hope.
However, replacing a simple PID loop or PDH with AI? Not sure if the extra complexity is needed; more so effort should be spent on making the technology cheaper and more accessible.
>However, replacing a simple PID loop or PDH with AI? Not sure if the extra >complexity is needed
Laser guy here. It's a highly-non-linear search problem (so not great for PID until you are close), and human intuition is used in surprising ways. For instance, maybe the operator knows that one adjuster is particularly 'wonky' and drifts more than the others; this is discovered through trial and error. Believe it or not, some people seem to have "the touch", and others never seem to get it (I think it might just be a matter of patience - alignment can be boring as hell and some people find it frustrating.) To be trained, usually someone starts out by "shadowing" a more experienced person, but eventually people develop their own style that can be hard to fully verbalize.
I am curious if they could de-tune the system, ask the AI to hit a target, and get anything sensible. Because the AI would be relentless, I bet it might work better than expected...
However, replacing a simple PID loop or PDH with AI? Not sure if the extra complexity is needed; more so effort should be spent on making the technology cheaper and more accessible.
Laser guy here. It's a highly-non-linear search problem (so not great for PID until you are close), and human intuition is used in surprising ways. For instance, maybe the operator knows that one adjuster is particularly 'wonky' and drifts more than the others; this is discovered through trial and error. Believe it or not, some people seem to have "the touch", and others never seem to get it (I think it might just be a matter of patience - alignment can be boring as hell and some people find it frustrating.) To be trained, usually someone starts out by "shadowing" a more experienced person, but eventually people develop their own style that can be hard to fully verbalize.
I am curious if they could de-tune the system, ask the AI to hit a target, and get anything sensible. Because the AI would be relentless, I bet it might work better than expected...
- This involve multiple variable, so it is never a simple PID loop
- They didn't use LLM for the feedback loop, they use the LLM to test and fine tune their program.