|
Sorry for the late reply, you probably won't see this but I had some work
troubles. My perspective on this is that Go (I don't know anything about rockets), was a
technological advance that should not have been surprising for anyone who knew
the state of the art. One reason I don't know how to interpret Coulom's quote is
that he must have been one of the few people who understood computer go as well. You'd say, yeah, that's the point, but if you look at the AlphaGo paper it
really has nothing new in it. There is no sudden breakthrough in scientific
understanding, in algorithm design, in efficiency or accuracy of a search, or
anything of the sort. All there is, is Monte Carlo Tree Search and neural nets.
That's as old as bread and about as exciting. The only reason why it was AlphaGo that dominated Go, and not some other system,
is that it was created by a company with the resources of DeepMind, meaning
Google. It had to be such a big company exactly because there was no real
advance in the scientific understanding of the game of Go, and so progress could
only be made by turning up the compute to 11. The same thing happenned back in the 1980's with DeepBlue and chess,
also: it had to be IBM because few others could make a minimax system that
searched so far and deep, because few others had IBM's computing resources. And you know, perhaps that's what Coulom meant when he said "10 years". That it
would take 10 years for the scientific understanding of Go to advance to the
point where a machine can beat a human. It just so happenned that turning on
the firehose and spamming the dollars at the problem made it go away, so there's
no need to understand anything anymore. Same thing happenned with chess, also. Alan Turing, Claude Shannon, John
McCarthy, all those AI greats, they thought that chess was a measure of human
intelligence and they wanted to understand how humans play chess using our human
intelligence. Well, it turned out that you don't need human intelligence to play
chess, as you don't need it to play Go, and a dumb machine with a blind search
can look deeper into a game tree and find better moves than a human. But as far
as I can tell, predictions of "10 years" for chess or for go or for anything
else, were trying to predict when the scientific understanding of those games
and of human and artificial intelligence would have advanced far enough that we
could make an intelligent machine capable of playing go like a human, but
better. That's what McCarthy would have wanted- he once made a bet with a
Grandmaster that "in 10 years" a machine would beat the grandmaster in chess;
and lost the bet. Because McCarthy was thiking about understaning intelligence,
not winning games. And he underestimated the difficult of the former and
overestimated that of the latter. Coulom, in the Wired quote, probably made the
same mistake, thinking along the same lines. If so, then he wasn't wrong, because while we have machines that can play better
Go than humans, we still don't understand how humans played better Go than
machines for so long. |