General Purpose AI
A while back, I wrote
about Google’s AI program called AlphaGo that beat the world human champion in
a 5 match series. The same company then created another AI program that was not
shown any games of Go masters to “learn” from. In fact, it was shown no games
at all! It was only fed the rules of the game, that’s all. And it had 4
specialized processors unlike the 48 of AlphaGo.
The speed at which
the new AI learnt Go was scorching. Within 3
days (yes, days), it was better than the AI that beat the human world
champion! Within 3 weeks, it could
60 of the world’s top players, including the updated version of AlphaGo!
Beating 60 top players is impressive because usually, such programs are custom
designed for one particular player and his style.
Then they trained
the AI’s guns on chess. Why chess, you wonder. Isn’t that old news in the man v/s
machine saga? Didn’t Deep Blue beat Garry Kasparov decades back? What’s left to
prove on the chess front?
Aha, unlike Deep
Blue, they used the same Go hardware for chess. Nor did they intend to pre-feed
it endless data of games played in the past or about chess openings. The AI
software called AlphaZero was only told the rules of chess. After some time of
learning, it played Stockfish,
one of the strongest (non-AI) computer programs there is. Further, Stockfish
ran on a system that was 900 times faster: that translates into an ability to
think 8 moves more.
The outcome of
this contest? AlphaZero crushed Stockfish: in a 100 game series, 28 wins, 72
draws, and 0 losses! How was this possible? The self-learning aspect aside, how
could a machine with 900 times less
processing power do this? Here’s Cornell University’s take on how it might have
happened:
“AlphaZero compensates for the lower number
of evaluations by using its deep neural network to focus much more selectively
on the most promising variations – arguably a more “human-like” approach to
search.”
Instead of a
semi-brute force approach, it was using a human like approach, except of
course, it was so much faster than any human at thinking like a human! And to
think it learnt all this from scratch, with no reference to how even openings
should be played.
What else could
AlphaZero do? Unlike conventional chess programs, it doesn’t abhor “blocked”
positions on the chessboard. It can create wins even from such positions. It
had even found solutions to unsolved problems!
And in case you
forgot, this was a generic system, not
something custom made for chess only. A generic AI can play out both ways: a
threat that takes over the world; or a boon that can evaluate and solve a wide
range of problems. Only the future can tell which way the AI chips will fall
for mankind…
Very interesting. The self-learning process, with the ability to apply beyond what the programmer told it to do by clear instructions, is pretty impressive.
ReplyDeleteI recall Kasparov's concluding remarks in his book that discussed the role he played in computerized chess. While he admitted that machines can, and will further, outperform humans is not in doubt, he only expressed a view that as for as feelings go, it is unlikely that robots and AIs are going to be our threatening competitors. We can't be sure of that, seeing the way all these developments happening. If feelings are due to the intelligence of the DNA (which has amazing potential seeing the vast number of species of differing "designs" - biologists will not agree to the design idea, I know - finally endowing mankind with intelligence which is used left, right and center too if polarization has not taken over the mind!), feelings can possibly be simulated by human programming too. And then request the AI to proceed further, hopefully avoiding the Frankenstein path!
There is another point to ponder:
The question that may not get addressed ever, possibly and hopefully, is whether the machines will outperform Newtions and Faradays and Einsteins and Eulers and Guasses; and of course Darwins and Michelangelos and Picassos and Shakespeares and Beethovens etc. Humankind has repeatedly shown the extraordinary inspiration, as we see in these people. We simply give a tag "genius" and may choose to express great admiration. Nevertheless we are yet to identify what goes into the making of the uncanny inspiration capacity. Not even an inkling so far about it, let alone hunch.
Will the AIs beat us to it too? AI geniuses may become dime a dozen, wow!! God only knows if it can happen, but as always, he is not going to tell us a thing! He always let's us do the mess-up! :-)