The Most Human Human (17 page)

And this is why Game 6 didn’t count. Kasparov bungled his seventh move (7 … h6, intending 8 … Bd6, instead of the correct 7 … Bd6 first, followed by 8 … h6), falling into a well-known book trap. The machine looked up the position and delivered the withering knight sacrifice (8.Nxe6) straight from its tables. Kasparov eventually wrenched himself out of book with a novel, if desperate, defense (11 … b5), but it was too late—by the time Deep Blue, that is, the search, analysis, and move selection procedure, finally stepped in, it was only to deliver the coup de grâce.

I agree with him that Game 6 “didn’t count.” He could have defended better, he could have held on longer, but at heart Kasparov lost that game
in book
.
¹¹
(One commenter online described it with the wonderful phrase “game six’s all-book crusher.”) Tripping and falling into a well on your way to the field of battle is not the same thing as dying on it.

And—and here’s a more metaphysical assertion, the one that takes us back to the Turing test and back to ourselves—whoever or whatever
achieved that winning position over Kasparov
was not Deep Blue
. “Hey! Who sacked my knight!?” Ashley jokingly imagines Deep Blue wondering, as its analysis function finally kicks in. Indeed.

Deep Blue is only itself out of book; prior to that it is nothing. Just the ghosts of the game itself.

And so it is with us
, I find myself saying.

Some of my friends growing up were competitive players of the game 24 in middle school. The idea of the game was that there were these cards of four numbers—say, 5, 5, 4, 1—and you had to figure out a way to add, subtract, multiply, or divide them in order to produce the number 24: in this case, 5 × (5 – 1)
+
4 would do it, for instance. New Jersey had an annual 24 tournament for the state’s best middle schoolers, and they got to go. Well, one of the kids there had spent basically the entire month before the finals simply
memorizing the cards
. He announced this to the other competitors at his table with a smirk. The finals were timed—first person to shout out the correct answer gets the point—and so the insinuation was that no one would have a chance: they’d be
calculating
, and he’d be
remembering
. When the official casually announced, in her opening remarks, that new cards had been prepared specially for this event, my friends watched, and may have suppressed a smirk or two themselves, as all the color drained from the kid’s face. The round began, and he got slaughtered.

Garry Kasparov observes that this memorization-based approach is alarmingly common among new players:

In
The Game
, Neil Strauss recounts trying to initiate a threesome with two women he’d recently befriended. Otherwise clueless, all he knows is what Mystery has told him he’s done in the past—run a bath and ask for help washing your back. So Strauss goes through the same moves, with predictably disastrous results:

Strauss finds himself in an awfully strange situation, of course, because he’s there in the tub … but he didn’t actually
want
a bath—so it’s the same “Who sacked my knight?” problem: “Who
asked for the back wash?” When the authentic Neil “kicks in,” how does he deal with the odd things he said and did when he was possessed by the book?

Neil Strauss argues that the nightlife of L.A.’s Sunset Boulevard has been blighted by a new generation of pickup artists—which he refers to as “social robots”—who have taken the art out of flirtation by replacing genuine conversational ability with mere and breathtakingly elaborate “opening line” repertoire.
¹²
“International dating coach” Vin DiCarlo is compiling a database of text messages, thousands upon thousands, and cataloging their success (“even down to the punctuation”) at prompting replies and dates. Increasingly-popular dating websites and books offer pre-scripted conversational openings, emphasizing memorization and repetition: “Once you have performed a particular story or routine dozens of times, you don’t even have to think about what you are saying. Your mind is free for other tasks, such as planning the next move. You have already fully explored all the conversational threads that could possibly arise from this piece of material. It’s almost like seeing the future.”

The fascinating thing is that we now have a
measure
for the failure of these conversational approaches. That measure is the Turing test—because its programmers are, ironically, using many of the selfsame approaches.

Not all conversational autopilot happens in courtship, though, and sometimes it overtakes those even with the best of intentions. News anchor and interviewer Ted Koppel bemoans, “It’s amazing how many
people come into an interview having already decided what their questions are going to be, having decided on the order of those questions, and then pay absolutely no attention to what the interviewee is saying. Frequently people reveal something about themselves in an interview, but if you don’t follow up on it, it will be lost.” Nor does this kind of unresponsive script-reciting and rule-following happen only to those who use it as a deliberate or half-deliberate strategy. I think we’ve all, at some moments or others, found ourselves running through the standard conversational patterns and “book” responses without realizing it, or have actively searched for ways to get a conversation out of book, but didn’t know how.

The history of AI provides us not only with a
metaphor
for this process but also with an actual
explanation
, even a set of benchmarks—and, better than that, it also suggests a solution.

Our first glimpse at what that solution might look like comes from the world of checkers, one of the first complex systems to be rendered “dead” by its book—and this was the better part of a century before the computer.

Checkers hit rock bottom in Glasgow, Scotland, in 1863.

Twenty-one games of the forty-game world-championship match then being played between James Wyllie and Robert Martins were
the exact same game
from start to finish. The other nineteen games, too, began with the same opening sequence, dubbed the “Glasgow opening,” and all forty were draws.

For checkers fans and organizers alike (you can only imagine the mind-blowingly insipid headlines this match must have generated, and how rankled the sponsors must have felt), the Wyllie-Martins 1863 match was the last straw. Opening theory, combined with top players’ take-no-risks attitude, had ground top-level checkers to a halt.

But what was to be done? How to save a dying game, rendered static by the accumulation and calcification of collective wisdom?
You couldn’t just
force
world-class checkers players to systematically not play the moves established as the correct ones—or could you?

Perhaps, if you didn’t like how players were opening their games, you could simply
open their games for them
. This is exactly what checkers’ ruling body began to do.

Starting in America around 1900, serious tournaments began operating with what’s called the “two-move restriction.” Before a match, the first two opening moves are chosen at random, and the players play two games from the resulting position, one from either side. This leads to more dynamic play, less reliance on the book, and—thank God—fewer draws. But after another generation of play, even the two-move restriction, with its 43 starting positions,
¹³
began to seem insufficient, and in 1934 it was upped to a three-move restriction with 156 different starting positions. Meanwhile, in an odd twist, classic checkers, with no move restrictions, has become itself a kind of
variant
, called “Go-As-You-Please.”
¹⁴

A new method of opening randomization, called “11-man ballot,” where one of the twelve pieces is removed at random from either side, and
then
the two-move restriction is applied, is now starting to gain traction. The number of starting positions in 11-man ballot checkers is in the thousands, and while the three-move restriction remains, since 1934, de rigueur at the top levels of play, it seems likely that the future of serious checkers, such as there is one, lies there.

Even when organizers aren’t forcing players to randomize their openings, there may be good strategic reasons to do so: one makes an admittedly, if only slightly, weaker move than the move opening theory prescribes, hoping to come out ahead by catching one’s opponent off guard. Garry Kasparov popularized this notion, dubbed
“anti-computer chess,” in his games against Deep Blue: “I decided to opt for unusual openings, which IBM would not have prepared for, hoping to compensate for my inferior position with superior intuition. I tried to present myself to the computer as a ‘random player’ with extremely odd playing characteristics.”
¹⁵

When computers play each
other
, the influence of the opening book is so dramatic and, frequently, decisive that the chess community began to distrust these games’ results. If you wanted to buy commercial chess software to analyze your own games and help you improve, how could you know which chess programs were stronger? The ones with the deepest books would dominate in computer tournaments, but wouldn’t necessarily be the best analytically. One answer, of course, would be simply to “unplug” the opening book from the analysis algorithm and have the two programs play each other while calculating from move one. But this, too, produces skewed results—picking good opening moves is a different beast from picking good middle- or endgame moves, and it’d be unfair, and irrelevant, to have programmers, for the purposes of winning these computer contests, spend weeks honing opening-move analysis algorithms when in practice (i.e., when the software has access to the opening book) this type of analysis will never be used.

It was English grandmaster John Nunn who first addressed this problem in the late 1990s by creating “test suites” of a half dozen or so unusual (that is, out of book), complex, and balanced middle-game positions, and having programs take turns playing from either side of each of these positions, for a total of a dozen games. The programs
simply begin playing “in medias res”—lopping off the opening phase of the game altogether.

In the beginning of the twenty-first century, former world champion Bobby Fischer shared these concerns, horrified at the generations of new players using computers to help them memorize thousands of book openings and managing to get the better of players with genuine analytical talent.
¹⁶
Chess had become too much about opening theory, he said, too much “memorization and prearrangement.” “The time when both players actually start thinking,” he said, “is being pushed further and further in.” He came to an even more dramatic conclusion than Kasparov and Nunn, however, concluding, “Chess is completely dead.”

His solution, though, was quite simple: scramble the order of the pieces in the starting position. Given a few basic guidelines and constraints (to maintain opposite-colored bishops and the ability to castle), you’re left with 960 different starting positions: enough to water down the opening book to near irrelevance. This version of the game
goes by “Fischer Random,” “Chess960,” or just “960.” Chess960 now has its own world championship, and many of the top world players of traditional chess also now play 960.

What of these rejuvenation efforts in checkers and chess? What might be the conversational equivalent? One is simply to be aware of what utterances occur frequently and attempt to step out of their shadow; for instance, “What’s your favorite color?” appears on over two million web pages, says Google, but “What’s your favorite brunch food?” appears on a mere four thousand. Correspondingly, Cleverbot has a very good answer for the former (“My favorite color is green”), but, lacking a specific brunch-related answer in its conversational book, must fall back on the more general “My favorite food is sea food.” It’s not a dead giveaway in a Turing test, but it’s a start, and more than enough to raise suspicion.

For the first few years, the Loebner Prize competition used specific topics of discussion, one for each program and confederate—things like Shakespeare, the differences between men and women, and the Boston Red Sox. The idea was to reduce the “domain” of discourse to a limited range, and give the computers a handicap while bot technology was still fledgling. Ironically, while removing the topic restriction, as they did in 1995, makes the computers’ task almost infinitely harder in
theory
, a glance at Loebner Prize transcripts suggests that it may have actually made their job
easier
in practice. Instead of gearing their software toward a specific topic area that changes each year, programmers spend year after year after year honing their software to start in the same way each time—with a friendly greeting and some small talk. I suspect the most difficult Turing test would be something akin to the 11-man ballot of checkers, or Chess960: a randomly assigned topic, chosen by the organizers just before the conversation begins. The computers that have regularly gotten the better of judges in recent years wouldn’t stand the whisper of a chance.