The fall of Fort Go

Another bastion of humanity has fallen. For decades, the ancient Chinese board game of go has been an outer marker for artificial intelligence research. It is played with much simpler rules than chess, but on a larger board, and is thereby much deeper mathematically. Go’s reputation amongst people who play both games well is that it is the more artistic, intuitive activity: it rewards thinking like a computer less well.

Some of this talk may just be the exoticizing charm of a pastime from the Far East. But go has really proved harder than chess for automata to master. Computer chess programs caught up with the best human players in the 1990s. Go, a game of “perfect information” like chess, was until recently thought to be a decade or more from that point. Progress toward the standard of the best humans had been frustrating.

But then, as sometimes happens in science, there came a s udden, jolting breakthrough. It didn’t involve advances in computing power or in library- searching techniques as such; at the basic level, the Googlesponsored AlphaGo program appears to use the same neural- network “learning machine” tricks that artificial intelligences have employed for a long time. The key seems to have been the introduction of a new division of “intellectual” labour into the programming — a committee of neural networks arrangement that works much better than the traditional ones.

Perhaps it is a more human arrangement, arrived at accidentally. (Is it obvious I don’t quite understand this scholarly paper?) At any rate, the Google team reported in a paper for Nature that its software was already able to beat other go programs 99.8 per cent of the time, and that it had licked the European go champion in a private match, 5- 0.

“European go champion” is a little like “best ice hockey player in Chile.” ( Seriously, white folks kinda stink at this game.) But AlphaGo is now playing a five- match series against Lee Sedol, a South Korean go professional of “nine- dan” rank, the highest attainable. He is generally considered the strongest player alive, with the emphasis now on that last word. On Thursday, Lee fell behind AlphaGo twonil, describing himself as “speechless” in a post- game press conference. An AP wire story described the champ as “grim and ashen” as he expressed hope that he might beat the computer at least once.

It is all exceedingly reminiscent of the famous 1997 match between Garry Kasparov, then the Lee Sedol of chess, and IBM’s Deep Blue supercomputer. Which, in turn, tells us how things are likely to pan out.

When he was narrowly beaten by Deep Blue, Kasparov behaved like the badtempered Russian genius he is, claiming that shady doings were afoot on the opposite side of the match. This might have seemed silly: whose advice were the programmers supposed to be using to beat Kasparov — God’s? But a great human chess player armed with a computer is stronger over the table than either the best computer or the best human. There was evidence that, with IBM’s reputation on the line, its chess consultants really were tweaking the software between games. After the match, Deep Blue was dismantled with suspicious alacrity.

The nastiness led to a sort of truce in chess that continues to this day. Computers are acknowledged to be better players than humans, because they have no nerve to lose and never visit the washroom at move 51. Computer evaluations of chess positions are broadly treated as objective truths, but not quite God’s word. Humans have found good ways to thwart computers in single matches by using specific tactics, usually bizarre-looking ones that would never occur in a human game.

Kasparov found clashing with the inhuman to be unfamiliar and exasperating. His skill at the “fight between two brains” turned out to rely, more than anyone thought, upon reading subtle physical cues in opponents. Sedol has now reported the same experience, though less irascibly: “Normally, you can sense your opponent’s breathing, their energy,” he complained to BBC News. The Korean is said to be an aggressive, unorthodox go player who ( like some top chess players) pays little heed to the canon of game openings. A slightly inferior player of different temperament might have more success against AlphaGo.

There is a lot of schoolmarmish terror of AIs going around, and the victory of Deep Blue was, at the time, widely felt to be ominous. But t he supremacy of chess AI has not ruined chess. It gives human learners and fans access to somewhat godlike evaluations of moves, and has enabled automated analysis of, and insight into, classic games. Professionals have to be more careful about cheating, but among elite players a computer’s chess moves played by a human swindler would be obvious. Computers turned out not to change the essential experience humans have when they sit down unassisted at a chessboard. Go will certainly be just the same. Cheer up, meatbags.

COMPUTERS STARTED BEATING HUMANS AT CHESS ALMOST 20 YEARS AGO. WE THOUGHT GO WOULD TAKE ANOTHER 10 YEARS. WE WERE WRONG.