Remi
Coulom (left) and his computer program, Crazy Stone, take on
grandmaster Norimoto Yoda in the game of Go. Photo:
Takashi Osato/WIRED
TOKYO,
JAPAN — Rémi Coulom is sitting in a rolling desk chair,
hunched over a battered Macbook laptop, hoping it will do something no
machine has ever done.
That
may take another ten years or so, but the long push starts here,
at Japan’s University of Electro-Communications. The venue is far from
glamorous — a dingy conference room with faux-wood paneling and garish
fluorescent lights — but there’s still a buzz about the place.
Spectators are gathered in front of an old projector screen in the
corner, and a ragged camera crew is preparing to broadcast the
tournament via online TV, complete with live analysis from two
professional commentators.
Coulom
is wearing the same turtleneck sweater and delicate rimless
glasses he wore at last year’s competition, and he’s seated next to his
latest opponent, an ex-pat named Simon Viennot who’s like a younger
version of himself — French, shy, and self-effacing. They aren’t
looking
at each other. They’re focused on the two computers in front of them.
Coulom’s is running a piece of software called Crazy Stone — the work
of
over seven years — and the other runs Nomitan, coded by Viennot and his
Japanese partner, Kokolo Ikeda.
Crazy
Stone and Nomitan are locked in a game of Go, the Eastern
version of chess. On each screen, you can see a Go board — a grid of 19
lines by 19 lines — filling up with black and white playing pieces,
each
placed at the intersection of two lines. If Crazy Stone can win and
advance to the finals, it will earn the right play one of the best
human
Go players in Japan. No machine has ever beaten a top human Go player —
at least not without a huge head-start. Even if it does advance to the
man-machine match, Crazy Stone has no chance of changing this, but
Coulom wants to see how far his creation has come.
Computers
match or surpass top humans in chess, Othello, Scrabble, backgammon,
poker, evenJeopardy.
But not Go. The
challenge is daunting. In 1994, machines took the checkers crown,
when a program called Chinook beat the top human. Then, three years
later, they topped the chess world, IBM’s Deep Blue supercomputer
besting world champion Garry Kasparov. Now, computers match or surpass
top humans in a wide variety of games: Othello, Scrabble, backgammon,
poker, even Jeopardy.
But not Go. It’s the one classic game where wetware still dominates
hardware.
Invented
over 2500 years ago in China, Go is a pastime beloved by
emperors and generals, intellectuals and child prodigies. Like chess,
it’s a deterministic perfect information game — a game where no
information is hidden from either player, and there are no built-in
elements of chance, such as dice.1
And like chess, it’s a
two-person war game. Play begins with an empty board, where players
alternate the placement of black and white stones, attempting to
surround territory while avoiding capture by the enemy. That may seem
simpler than chess, but it’s not. When Deep Blue was busy beating
Kasparov, the best Go programs couldn’t even challenge a decent
amateur.
And despite huge computing advances in the years since — Kasparov would
probably lose to your home computer — the automation of expert-level Go
remains one of AI’s greatest unsolved riddles.
Rémi
Coulum is part of a small community of computer scientists
hoping to solve this riddle. Every March, the world’s most dedicated Go
programmers gather at the University of Electro-Communications to
compete in the UEC Cup, a computer Go tournament that, uniquely,
rewards
two finalists with matches against a “Go sage,” the equivalent of a
chess grandmaster. Organizers dub these machine-versus-man matches the
Densei-sen, or “Electric Sage Battle.”
At
this year’s UEC Cup, Coulom’s Crazy Stone is the favorite. On the
first day of the competition, the software program went undefeated,
which earned it top seed in today’s 16-member single-elimination
bracket
and a bye in the first round. Now, it’s the second round, and Viennot,
a
relative newcomer to the computer Go scene, tells me he’ll be happy if
his program just puts up a good fight. “Nomitan uses many of Rémi’s
tricks, but I don’t think it will be enough,” he says. “Crazy Stone is
a
much stronger program.”
Rémi
Coulom and Crazy Stone. Photo:
Takashi Osato/WIRED Even
for Coulom — a good but not great Go player himself — Crazy Stone’s
moves can be incomprehensible.
THe
computer screens in front of Coulom and Viennot display
statistics that show the relative confidence of each program. Although
the match has just begun, Crazy Stone is already 58 percent sure it
will
prevail. Oddly, Nomitan’s confidence level is about the same. When I
point this out to Coulom and Viennot, they both laugh. “You can’t trust
these algorithms completely,” explains Viennot. “They are always a
little over-confident.”
The
official commentary doesn’t start until the final match, but as
the second round progresses, a small crowd forms around commentator
Michael Redmond to hear his thoughts. The charismatic Redmond, an
American, is one of very few non-Asian Go celebrities. He began playing
professionally in Japan at the age of 18, and remains the only
Westerner
to ever reach 9-dan, the game’s highest rank. “I don’t know the black
player,” he says, referring to Nomitan, “but it has a flashy style,
flashier than Crazy Stone. Very good tesuji. With humans, tesuji are a
fairly accurate gauge of strength, and now, I’m seeing computers do
them
more.”
Tesuji
means something like “clever play,” and Nomitan’s tesuji are
giving Crazy Stone serious trouble. With the game nearly halfway done,
Crazy Stone is only 55 percent confident, which means it’s even money.
After a few more turns, another professional named O Meien pronounces
Nomitan the leader. As other games in the room finish, the crowd in
front of the projector screen grows larger and louder. From the sound
of
it, Crazy Stone’s prospects are increasingly bleak.
Most
people in the room take the pros like O Meien at their word. We
have to, since games of Go are often so complex that only extremely
high-level players can understand how they’re progressing. Even for
Coulom — a good but not great Go player himself — Crazy Stone’s moves
can be incomprehensible. But Coulom identifies as a programmer more
than
a player, which allows him to remain calm in the face of professional
skepticism. He trusts the confidence level Crazy Stone shows him.
“Maybe
O Meien is thinking about which side looks better,” he says, with a
lilting French accent. “But I know Crazy Stone is much stronger than
Nomitan. So I just think at some point Nomitan will probably mess up.”
And
so it does. Crazy Stone makes a number of moves that prompt
murmurs of approval from the crowd. Despite those initial tesuji,
Nomitan squanders its advantage. Soon, Crazy Stone’s confidence levels
are in the high 80s, and Nomitan resigns.
The
other matches leading up to the final are uneventful, with the
exception of one semi-final contest. Zen, Crazy Stone’s biggest rival
and last year’s runner-up, nearly loses to a program called Aya. The
game begins with a complicated local battle in the upper right corner,
each side trying to keep their stones alive. At first, Zen plays with
excellent kiai, or fighting spirit. The area looks settled. Then,
without warning, Zen makes an obvious mistake, eliciting a collective
gasp from the room. Zen’s co-programmer, a Japanese man with long
graying hair named Hideki Kato, keeps his eyes on the confidence levels
streaming across his laptop screen, and eventually, Zen manages to eke
out a lead, before Aya resigns. The final is decided, a rematch of last
year’s match: Crazy Stone vs. Zen.
The
Mystery of Go
Even
in the West, Go has long been a favorite game of mathematicians,
physicists, and computer scientists. Einstein played Go during his time
at Princeton, as did mathematician John Nash. Seminal computer
scientist Alan Turing was a Go aficionado, and while working as a World
War II code-breaker, he introduced the game to fellow cryptologist I.J.
Good. Now known for contributing the idea of an “intelligence
exposition” to singularity
theories
— predictions of how machines will become smarter than people — Good
gave the game a huge boost in Europe with a 1965 article for New
Scientist entitled “The Mystery of Go.”
A
woman and a man play Go in Korea sometime in the early 1900s. Photo:
Library
of Congress
Good
opens the article by suggesting that Go is inherently superior
to all other strategy games, an opinion shared by pretty much every Go
player I’ve met. “There is chess in the western world, but Go is
incomparably more subtle and intellectual,” says South Korean Lee
Sedol,
perhaps the greatest living Go player and one of a handful who make
over seven figures a year in prize money. Subtlety, of course, is
subjective. But the fact is that of all the world’s deterministic
perfect information games — tic-tac-toe, chess, checkers, Othello,
xiangqi, shogi — Go is the only one in which computers don’t stand a
chance against humans.
‘There
is chess in the western world, but Go is incomparably more subtle and
intellectual.’
This
is not for lack of trying on the part of programmers, who have
worked on Go alongside chess for the last fifty years, with
substantially less success. The first chess programs were written in
the
early fifties, one by Turing himself. By the 1970s, they were quite
good. But as late as 1962, despite the game’s popularity among
programmers, only two people had succeeded at publishing Go programs,
neither of which was implemented or tested against humans.
Finally,
in 1968, computer game theory genius Alfred Zobrist authored
the first Go program capable of beating an absolute beginner. It was a
promising first step, but notwithstanding enormous amounts of time,
effort, brilliance, and quantum leaps in processing power, programs
remained incapable of beating accomplished amateurs for the next four
decades.
To
understand this, think about Go in relation to chess. At the
beginning of a chess game, White has twenty possible moves. After that,
Black also has twenty possible moves. Once both sides have played,
there
are 400 possible board positions. Go, by contrast, begins with an empty
board, where Black has 361 possible opening moves, one at every
intersection of the 19 by 19 grid. White can follow with 360 moves.
That
makes for 129,960 possible board positions after just the first round
of moves.
The
rate at which possible positions increase is directly related to a
game’s “branching factor,” or the average number of moves available on
any given turn. Chess’s branching factor is 35. Go’s is 250. Games with
high branching factors make classic search algorithms like minimax
extremely costly. Minimax creates a search tree that evaluates possible
moves by simulating all possible games that might follow, and then it
chooses the move that minimizes the opponent’s best-case scenario.
Improvements on the algorithm — such as alpha-beta
search and null-move
— can prune the chess game tree, identifying which moves deserve more
attention and facilitating faster and deeper searches. But what works
for chess — and checkers and Othello — does not work for Go.
‘I’ll
see a move and be sure it’s the right one, but won’t be able to tell
you exactly how I know. I just see it.’
The
trouble is that identifying Go moves that deserve attention is
often a mysterious process. “You’ll be looking at the board and just
know,” Redmond told me, as we stood in front of the projector screen
watching Crazy Stone take back Nomitan’s initial lead. “It’s something
subconscious, that you train through years and years of playing. I’ll
see a move and be sure it’s the right one, but won’t be able to tell
you
exactly how I know. I just see it.”
Similarly
inscrutable is the process of evaluating a particular board
configuration. In chess, there are some obvious rules. If, ten moves
down the line, one side is missing a knight and the other isn’t,
generally it’s clear who’s ahead. Not so in Go, where there’s no easy
way to prove why Black’s moyo is large but vulnerable, and White has
bad
aji. Such things may be obvious to an expert player, but without a good
way to quantify them, they will be invisible to computers. And if
there’s no good way to evaluate intermediate game positions, an
alpha-beta algorithm that engages in global board searches has no way
of
deciding which move leads to the best outcome.
Not
that it matters: Go’s impossibly high branching factor and state
space (the number of possible board configurations) render full-board
alpha-beta searches all but useless, even after implementing clever
refinements. Factor in the average length of a game — chess is around
40
turns, Go is 200 — and computer Go starts to look like a fool’s errand.
A
traditional Go gameboard. Photo:
Takashi Osato/WIRED
Nonetheless,
after Zobrist, Go programmers persisted in their efforts
and managed to make incremental progress. But it wasn’t until 1979 that
a five-year project by computer scientist Bruce Wilcox produced a
program capable of beating low-level amateurs. As a graduate student at
the University of Michigan, Wilcox and his advisor collected detailed
protocols from games played against James Kerwin, who soon after would
leave for Japan to become the second-ever Western professional Go
player.
Unlike
successful chess programmers, Wilcox focused almost entirely
on modeling expert intelligence, collecting a vast database of stone
relationships from Kerwin’s games. His program divided the board into
smaller, more manageable zones, and then used the database to generate
possible moves, applying a hierarchal function to choose the best among
them. Forward-looking searches like alpha-beta, long the cornerstone of
AI gaming, were entirely absent from the program’s first incarnation
Then,
somewhat abruptly, progress stalled. The programs had encountered an
obstacle that also gives human players trouble.
During
the development process, Wilcox became a very strong amateur
player, an indispensable asset for early Go programmers, given that
programs depended so much on a nuanced understanding of the game. Mark
Boon (Goliath), David Fotland (Many Faces of Go), Chen Zhixing
(Handtalk
and Goemate) — the winners of computer Go competitions throughout the
80s and 90s — were all excellent players, and it was their combined
prowess as players and programmers that facilitated steady improvements
through the 90s. Then, somewhat abruptly, progress stalled. The
programs
had encountered an obstacle that also gives human players trouble.
“A
lot of people peak out at a certain level of amateur and never get
any stronger,” David Fotland explains. Fotland, an early computer Go
innovator, also worked as chief engineer of Hewlett Packard’s PA-RISC
processor in the 70s, and tested the system with his Go program.
“There’s some kind of mental leap that has to happen to get you past
that block, and the programs ran into the same issue. The issue is
being
able to look at the whole board, not the just the local fights.”
Fotland
and others tried to figure out how to modify their programs
to integrate full-board searches. They met with some limited success,
but by 2004, progress stalled again, and available options seemed
exhausted. Increased processing power was moot. To run searches even
one
move deeper would require an impossibly fast machine. The most
difficult game looked as if it couldn’t be won.
Enter
Rémi Coulom, whose Crazy Stone would inaugurate a new era of
computer Go. Coulom’s father was a programmer, and in 1983, he gave his
son a Videopac computer for Christmas. Coulom was nine, around the time
most Go prodigies leave home to begin intensive study at an academy.
After less than a year, he had programmed Mastermind. In four years, he
had created an AI that could play Connect Four. Othello followed
shortly
thereafter, and by 18, Coulom had written his first chess program.
Enter
Rémi Coulom, whose Crazy Stone would inaugurate a new era of computer
Go.
The
program, Crazy Bishop, was awful. Without access to the internet,
Coulom had to invent everything from scratch. But a year later, he
started engineering school, where university computers allowed him to
swap algorithms and strategies in online chess programming communities.
Crazy Bishop improved quickly. In 1997, the year Deep Blue defeated
Kasparov, Coulom attended the world computer chess championship in
Paris, where he made a decent showing and met members of his online
community in person. The event inspired him to continue graduate study
as a programmer, not an engineer. Following a stint in the military and
a
masters in cognitive science, Coulom earned a PhD for work on how
neural networks and reinforcement learning can be used to train
simulated robots to swim.
Although
he’d encountered Go at the 2002 Computer Olympiad, Coulom
didn’t give it much thought until 2005, when, after landing a job at
the
University of Lille 3, he began advising Guillaume Chaslot, a masters
student who wanted to write a computer Go program as his thesis.
Chaslot
soon left to start his PhD, but Coulom was hooked, and Go became a
full-time obsession.
The
Monte Carlo Bet
It
wasn’t long before he made his breakthrough. Coulom had exchanged
ideas with a fellow academic named Bruno Bouzy, who believed that the
secret to computer Go might lie in a search algorithm known as Monte
Carlo. Developed in 1950 to model nuclear explosions, Monte Carlo
replaces an exhaustive search with a statistical sampling of fewer
possibilities. The approach made sense for Go. Rather than having to
search every branch of the game tree, Monte Carlo would play out a
series of random games from each possible move, and then deduce the
value of the move from an analysis of the results.
Rather
than having to search every branch of the game tree, Monte Carlo would
play out a series of random games from each possible move. Bouzy
couldn’t make it work. But Coulom hit upon a novel way of
combining the virtues of tree search with the efficiency of Monte
Carlo.
He christened the new algorithm Monte Carlo Tree Search, or MCTS, and
in January of 2006, Crazy Stone won its first tournament. After he
published his findings, other programmers quickly integrated MCTS into
their Go programs, and for the next two years, Coulom vied for
dominance
with another French program, Mogo, that ran a refined version of the
algorithm.
Although
Crazy Stone ended up winning the UEC Cup in 2007 and 2008,
Mogo’s team used man-machine matches to win the publicity war. Coulom
felt the lack of attention acutely. When neither the public nor his
university gave him the recognition he deserved, he lost motivation and
stopped working on Go for nearly two years.
Coulom
might have given up forever had it not been for a 2010 email
from Ikeda Osamu, the CEO of Unbalance, a Japanese computer game
company. Ikeda wanted to know if he’d be willing to license Crazy
Stone.
Unbalance controlled about a third of the million-dollar global market
in computer Go, but Zen’s commercial version had begun to increase its
market share. Ikeda needed Coulom to give his company’s software a
boost.
The
first commercial version of Crazy Stone hit the market in spring
of 2011. In March of 2013, Coulom’s creation returned to the UEC Cup,
beating Zen in the finals and — given a four-stone head-start — winning
the first Densei-sen against Japanese professional Yoshio “The
Computer”
Ishida. The victories were huge for Coulom, both emotionally and
financially. You can see their significance in the gift shop of the
Japan Go Association, where a newspaper clipping, taped to the wall
behind display copies of Crazy Stone, shows the pro grimly succumbing
to
Coulom’s creation.
Photo:
Takashi Photo:
Takashi Osato/WIRED
Extremely
Human
During
the break before this year’s UEC final, the TV crew springs
into action, setting up cameras and adjusting boom mikes. Redmond,
microphone in hand, positions himself at the front of the room next to
the magnetic board. On the other side is Narumi Osawa, a pixieish 4-dan
professional who, in standard Japanese fashion, will act as an
obsequious female foil — “What was that? Oh, wow, I see! Hai! Hai!” —
for Redmond’s in-game analysis.
Once
everything is in place, Kato and Coulom are called to the front
of the room for nigiri, to determine who plays first. Since he is the
favorite, Coulom reaches into one of two polished wooden goke and grabs
a
fistful of white stones. Kato places one black stone on the board,
indicating his guess that Coulom holds an odd number of stones. The
white stones are counted. Kato guessed correctly. He will be Black, and
the game is underway.
The
move is utterly bizarre, and even Kato is somewhat baffled.
It
takes only three turns before the room explodes with excitement.
After claiming two star points in the corners — a standard opening —
Zen
has placed its third stone right near the center of the board. The move
is utterly bizarre, and even Kato is somewhat baffled. “An inhuman
decision,” Viennot whispers to me. “But Zen likes to make moyo in the
middle of the board, like Takemiya. Maybe this is a new style.”
Kato
and Coulom are sitting next to each other, eyes fixed on their
laptops, occasionally exchanging confidence levels. An interesting
struggle develops in the upper left corner, where Crazy Stone has
invaded and Zen is trying to strengthen its position. The crowd mutters
when Redmond pronounces one of Zen’s moves “extremely human.” (“Hai!
Hai!”) Black and white stones continue to fill the board, beautiful as
always, forming what is technically known as a percolated fractal.
Suddenly,
Coulom tenses up. Crazy Stone’s confidence levels are
rising quickly, too quickly, and soon, they are far too high, up in the
sixties. It appears the program has misjudged a semeai, or capturing
race, and believes a group of stones in the upper right corner is safe,
when in fact it is not. Since Crazy Stone’s move choices depend on an
accurate assessment of the overall board position, the misjudged group
proves fatal. On its 186th move, Crazy Stone resigns, and Zen becomes
the new UEC Cup champion.
Later
that evening, at the celebratory banquet, Coulom says he
doesn’t feel too bad, but I suspect he’s extremely disappointed. Still,
there’s a chance for redemption. As a finalist, Crazy Stone gets to
compete in the Densei-sen.
The
Electric Sage Battle
Coulom
plays down the Electric Sage Battle. “The real
competition is program against program,” he told me during one early
phone interview. “When my opponent is a programmer, we are doing the
same thing. We can talk to each other. But when I play against a
professional and he explains the moves to me, it is too high level. I
can’t understand, and he can’t understand what I am doing. The
Densei-sen — it is good for publicity. I am not so interested in that.”
But
when we meet at the Densei-sen, he seems excited. The building is
humming with activity. Last weekend’s conference room is reserved for
press and university dignitaries, and a new, private room has been
equipped for the matches. Only the referee and timekeepers will be
allowed in the room, and cameras have been set up to capture the action
for the rest of us. The professional commentators are now in the
building’s main auditorium, where at least a hundred people and three
TV
crews are ready to watch Crazy Stone and Zen take on a real pro.
In
2013, the Electric Sage Battle starred Ishida “The Computer”
Yoshio, so-called because of his extraordinary counting and endgame
abilities. This year, the pro is Norimoto Yoda, known for leading the
Japanese team to a historic victory over Korea in the 2006 Nongshim
Cup,
and for shattering Go stones when he slams them down on the hardwood
goban. After an introductory ceremony, Coulom and Yoda enter the
private
room, bow, and take their seats. In his typical style, Yoda has come
dressed in an olive green kimono. His left hand holds a folded fan.
Coulom, in his typical style, is wearing a blue turtleneck sweater. On
the wooden goban between them sit two gokes filled with stones — Black
for Coulom, White for Yoda.
In
his typical style, Yoda has come dressed in an olive green kimono. His
left hand holds a folded fan
This
time, there is no nigiri. Crazy Stone receives a massive
handicap, starting with four black stones placed advantageously on the
corner star points (the 4 by 4 intersections on a Go board’s 19 by 19
grid). Yoda has no choice but to adopt an aggressive style of play,
invading Crazy Stone’s territory in hopes of neutralizing his initial
disadvantage. But Crazy Stone responds skillfully to every threat, and
Yoda’s squarish face starts to harden. The fan snaps open and shut,
open
and shut.
In
the press room, we can’t hear the auditorium commentary. Instead, I
watch as Muramatsu Murakasu, a main organizer of the event, plays the
game out on his own board with O Meien. The two take turns trying to
predict where Yoda and Crazy Stone will move next, and as the game
progresses, both agree that Crazy Stone is doing an excellent job
maintaining its lead.
Meanwhile,
Coulom is looking at the board, his laptop, the
timekeepers, anywhere but the increasingly frustrated Yoda. After
Coulom
places one particular stone, Yoda’s eyes narrow perceptibly. He grunts
and fans himself furiously. “That was an excellent move,” says O Meien.
“Yoda-san must be upset.”
Crazy
Stone continues to play brilliant Go, and all of Yoda’s
incursions prove fruitless. It is only as the end approaches that Crazy
Stone reveals its true identity. With a lead of eleven points, any
decent human in Crazy Stone’s position would play a few obvious moves
and then pass, allowing Yoda resign. But Crazy Stone’s algorithm is
structured to care only about winning — not by how much. Coulom winces
as Crazy Stone makes a wasted move in its own territory, and then
another. The game drags on as Crazy Stone sacrifices points, until
mercifully it decides to pass, and the machine is finally declared the
winner.
Coulom
leaves the fuming Yoda as quickly as possible and joins us in
the press room. He’s both ecstatic and mortified. “I am proud of Crazy
Stone,” he says. “Very proud. But the first thing I will do at home is
work on the endgame, so it does not make such embarrassing moves.” Then
things get better. Yoda manages to beat Zen in the second Densei-sen
match, and just like that, the glory of the Electric Sage Battle
belongs
to Coulom, whose program has now bested two professionals after a
four-stone handicap.
Photo:
Takashi Osato/WIRED
When
AI Is Not AI
After
the match, I ask Coulom when a machine will win without a
handicap. “I think maybe ten years,” he says. “But I do not like to
make
predictions.” His caveat is a wise one. In 2007, Deep Blue’s chief
engineer, Feng-Hsiung Hsu, said much the same thing. Hsu also favored
alpha-beta search over Monte Carlo techniques in Go programs,
speculating that the latter “won’t play a significant role in creating
a
machine that can top the best human players.”
Even
with Monte Carlo, another ten years may prove too optimistic.
And while programmers are virtually unanimous in saying computers will
eventually top the humans, many in the Go community are skeptical. “The
question of whether they’ll get there is an open one,” says Will
Lockhart, director of the Go documentary The
Surrounding Game. “Those who are familiar with just how strong
professionals really are, they’re not so sure.”
According
to University of Sydney cognitive scientist and complex
systems theorist Michael Harré, professional Go players behave in ways
that are incredibly hard to predict. In a recent study, Harré analyzed
Go players of various strengths, focusing on the predictability of
their
moves given a specific local configuration of stones. “The result was
totally unexpected,” he says. “Moves became steadily more predictable
until players reached near-professional level. But at that point, moves
started getting less predictable, and we don’t know why. Our best guess
is that information from the rest of the board started influencing
decision-making in a unique way.”
‘Moves
became steadily more predictable until players reached
near-professional level. But at that point, moves started getting less
predictable, and we don’t know why.’
This
could mean that computer programs will eventually hit another
wall. It may turn out that the lack of progress experienced by Go
programs in the last year is evidence of yet another qualitative
division, the same one that divides amateurs from professionals. Should
that be the case, another breakthrough on the level of the Monte Carlo
Tree Search could be necessary before programs can challenge pros.
I
was surprised to hear from programmers that the eventual success of
these programs will have little to do with increased processing power.
It is still the case that a Go program’s performance depends almost
entirely on the quality of its code. Processing power helps some, but
it
can only get you so far. Indeed, the UEC lets competitors use any kind
of system, and although some opt for 2048-processor-core
super-computers, Crazy Stone and Zen work their magic on commercially
available 64-core hardware.
Even
more surprising was that no programmers think of their creations
as “intelligent.” “The game of Go is spectacularly challenging,” says
Coulom, “but there is nothing to do with making a human intelligence.”
In other words, Watson and Crazy Stone are not beings. They are
solutions to specific problems. That’s why its inaccurate to say that
IBM Watson will be used to fight cancer, unless playing Jeopardy
helps reduce tumors. Developing Watson might have led to insights that
help create an artificial diagnostician, but that diagnostician isn’t
Watson, just as MCTS programs used in hospital planning are not Crazy
Stone.
The
public relations folks at IBM paint a different picture, and so
does the press. Anthropomorphized algorithms make for a better story.
Deep Blue and Watson can be pitted against humans in highly produced
man-machine battles, and IBM becomes the gatekeeper of a new era in
artificial intelligence. Caught between atheism and a crippling fear of
death, Ray Kurzweil and other futurists feed this mischaracterization
by
trumpeting the impending technological apotheosis of humanity, their
breathless idiocy echoing through popular media. “The Brain’s Last
Stand,” read the cover of Newsweek
after Kasparov’s defeat. But
in truth, these machines are nowhere close to mimicking the brain, and
their creators admit as much.
Many
Go players see the game as the final bastion of human dominance
over computers. This view, which tacitly accepts the existence of a
battle of intellects between humans and machines, is deeply misguided.
In fact, computers can’t “win” at anything, not until they can
experience real joy in victory and sadness in defeat, a programming
challenge that makes Go look like tic-tac-toe. Computer Go matches
aren’t the brain’s last stand. Rather, they help show just how far
machines have to go before achieving something akin to true human
intelligence. Until that day comes, perhaps it’s best to view the
Densei-sen as programmers do. “It is fun for me,” says Coulom, “but
that’s all.”
Alan
Levinovitz is assistant professor of philosophy and religion
at James Madison University. His writing has appeared in Slate, Salon,
The Believer, and elsewhere. Follow him: @top_philosopher.
1Update
00:05 EST 05/13/14: An earlier version of this
story referred to Go as a “perfect information game.” It is more
accurate to call it a “deterministic perfect information game.”
Photo:
Takashi Photo:
Takashi Osato/WIRED