2 ms·
> I have always thought he was too quick to dismiss the potential of AI I have no idea what AI is (in a practical context), and I think no one else does, eithe
by pron 10y ago
> I have always thought he was too quick to dismiss the potential of AI
I have no idea what AI is (in a practical context), and I think no one else does, either. Do you mean the statistical algorithms that people call machine learning?
> But we now live in a world where, at long last, a machine has beaten the world champion Go player.
It is very important to understand how that was achieved and the difference between that and verification. It's unknown how good Google's Go program is compared to some arbitrary opponent, as the algorithm it employs is not the algorithm to "win" at Go (which is far too computationally expensive to be feasible). It is "merely" able to beat other human players, but that is mostly because humans can beat human players, and the program learned from humans. Humans, however, do not efficiently verify other humans' programs, and that makes the problem very different.
Don't get me wrong: I think that statistics are absolutely necessary, and that machine learning could certainly help. But "help" does not necessarily mean "solve". At the end of the day, even if we pretend machine learning does resemble intelligence (which it does not, as any expert would tell you; we're still in the vicinity of zero IQ), we must remember that intelligence -- unlike what some people like to believe -- is not a very good "general problem solving" algorithm; it is a reasonable general algorithm for problems that humans tend to confront in many situations. For example, humans are even worse than computers in finding solutions to NP complete problems, and are in general terrible at solving many problems. In general, humans are good at forming conceptually deep but undetailed theories. Is verifying programs one of those problems that intelligence can solve? It could be, but we have no good theory that can give us intuition into whether it is or it isn't. I do know that a model checker was far better than me when reasoning about programs.
But instead of hypothesizing, a good first step would be to gather statistics, and to use machine learning in analyzing those statistics. I think it's more likely that machine learning would uncover important patterns that we could then exploit in more robust algorithms.
Thanks for the invitation, but I am very far away from Silicon Valley.