3 ms·
I insisted to me Go-playing friend a number of years ago that what we need is an AI that can teach, as in, can you explain why this choice was made and why thes
by hinkley 4y ago
I insisted to me Go-playing friend a number of years ago that what we need is an AI that can teach, as in, can you explain why this choice was made and why these others were bad?
I'm not sure what that would exactly look like, but one of the things I like about Monte Carlo is that it works much more like how humans actually behave, and in theory when pressed on 'why', it could also behave more like humans behave: When challenged you (or at least I) run a more detailed simulation to pick some particularly juicy counter-examples. And if anywhere in there I sense self-doubt (why did I make that choice? Is it sound?) that may trigger a more exhaustive simulation. Turns out my pupil was right/wrong and now I understand the problem better myself.
What bothers me with respect to AI and patents is that if you did this, then you basically have a human rubber stamping an AI design. Eventually you'll end up with more subtle systems that resemble police work: this evidence is inadmissible in court but it does eliminate a set of suspects, so now I need to build a chain of clean evidence that points to an arrest. The latter might involve some skill and sophistication on the part of the human. The former just needs someone who can pass a sniff test on materials science or mechanical engineering who's willing to lie for money.