Une lecture en 23 passages.
Instead of a computer that thought and played chess like a human, with human creativity and intuition, they got one that played like a machine, systematically evaluating up to 200 million possible moves on the chess board per second and winning with brute number-crunching force.
This isn't to diminish the achievement in any way. It was a human achievement, after all, so while a human lost the match, humans also won.
Reexamining every aspect of that infamous match with Deep Blue for the first time in twenty years has been difficult, I admit. For two decades I have succeeded almost completely in avoiding and deflecting discussion about my Deep Blue matches beyond what was publicly known. There are many books about Deep Blue, but this is the first one that has all the facts and the only one that has my side of the story.
This sort of thinking is a trap into which every generation falls wher it comes to machine intelligence. We confuse performance—the abil ity of a machine to replicate or surpass the results of a human—with method, how those results are achieved. This fallacy has proved irre- sistible in the domain of higher intelligence that is unique to Homo sapiens.
Machines don't need to do things the same way the natural world does in order to be useful, or to surpass nature. This is obvious from millennia of physical technology and it applies to software and artifi. ciallv intelligent machines as well. Airplanes don't flap their wings and helicopters don't need wings at all. The wheel doesn't exist in nature but it has served us verv well. So whv should computer brains work like human brains in order to achieve results? As is so often the case in the crossroads of human and machine thinking, chess proved to be an ideal laboratory for investigating this question.
We don't get to pick and choose when technological progress stops or where. Companies are globalized and labor is becoming nearly as fluid as capital.
Few people in the world know better than I do what it's like to have your life's work threatened by a machine. No one was sure what would happen if and when a chess machine beat the world champion. Would there still be professional chess tournaments? Would there be sponsorship and media coverage of my world championship matches if people thought the best chess player in the world was a machine? Would people still play chess at all?
The existence of a Grandmaster-level chess program available on an inexpensive personal computer upended that hierarchy. While not as good as an experienced human coach, it was far better than nothing combined with the Internet's ability to bring the game to every corner of the world, a shift was under way.
The kev factor in producing elite chess talent is finding it earlv, and thanks to strong computers this is now very easv to do iust about anvwhere.
It's no coincidence that the current list of elite chess players contains many representatives of countries with little or no old chess traditions. Computers tend to have this impact in many ways, reducing the influence of dogma.
It took thirty years, but my beloved game was revealed to be too vulnerable to brute force fast searching to require strategic thinking from machines in order to defeat the best humans.
As much work as went into tuning Deep Blue's evaluation function and training its openings the depressing truth is that a few vears and a new generation of faster chips later, none of it would have mattered verv much.
For better or worse, chess just wasn't deep enough to force the chess-machine com- munity to find a solution beyond speed, something many among them lamented.
John McCarthy, the American computer scientist who coined the term "artificial intelligence" in 1956, called chess the "Drosophila of AI,' referring to how the humble fruit flv was the ideal subiect for countless seminal scientific experiments in biology, especially genetics.
By the end of the 1980s, the computer chess community had largely resigned this great experiment. In 1990, Ken Thompson of Belle was openly recommending the game Go as a more promising target for real advances in machine cognition.
In the same year, the compendium Computers, Chess, and Cognition. included an entire section on Go. titled "A New Drosophila for AI?"
The brute force "exhaustive search" method of checking every one of millions of positions also meant that the search tree couldn't reach very deep. If you could find a tactical threat that struck the de cisive blow four moves (eight ply) away when the computer could onl see three moves (six ply) deep, it wouldn't see it coming until it was too late. We call this the "horizon effect," exploiting that the machine can't see beyond its search "horizon."
Right around when Deep Thought became the first chess machine to become a real threat to Grandmasters at the end of the 1980s, artifi- cial intelligence was experiencing a broad resurgence in the scientific and business worlds.
The so-called AI winter that had descended after vears of overpromising and subsequent disillusionment was lifting. The crisis for AI stemmed from the evaporation of the confidence sa many experts in the 1970s had had in quickly discovering the secrets of cognition.
Research proiects and commercial AI ventures were shut- tered throughout the 198os and the AI movement had splintered. Basic science was out, practical systems were in. Understanding human in- telligence was passé, getting results in a narrow field was the fashion. The maxim became "Don't make it think, just make it work."
To be fair, those AI pioneers took aim at the biggest targets, like using natural language, self-teaching machines, and understanding abstract concepts. Still, their optimism sounds wildly out of proportion in hindsight. The 1956 Dartmouth Summer Research Project that launched the field of AI boldly proclaimed that great progress would be made on all of these things "if a carefully selected group of scientists work on it together for a summer." A summer!
The current world champion, Magnus Carlsen, is from Norway and was born in 1990. He has never known a world in which computer chess programs weren't stronger than he is. Ironically, he is very much a "human style" player, whose intuitive positional chess does not directly reflect much silicon influence. This is not the case for many of his contemporaries, however, something we will examine more closely later on.
Thanks to the international rating system we use in chess to rank players, a simple chart can show us quite clearly that chess machines have gotten stronger on a steady linear path from the first mainframes to specialized hardware machines to the top programs today.
They went from novice level in the 1960S to strong play in the seventies to Grandmaster level in the late eighties and world champion level in the late nineties there were no giant leaps, just a slow and steady evolution as the global community of developers learned from each other and competed with each other while Moore's law worked its inexorable magic on their hardware.
The apparent similarity between early chess programs and humar beginners is a trap, part of the familiar fallacy of expecting comput ers to think like humans. As Moravec's paradox dictates, computers are very good at chess calculation, which is the part humans have the most trouble with.
Computers are poor at recognizing patterns and making analogical evaluations, a human strength. Other than check mate, nearly every factor that goes into evaluating a chess position is conditional on many other factors. This, along with the slow speed of computers at the time, is why early experts thought it would be impos- sible to make a strong Type A (brute force) program.
They were wrong, although it would take a while to figure that out Many of the first programs were attempts at Type B, which sought to intelligently reduce the size of the algorith m's search tree early on the way humans do. Other research groups saw the advantage of tackling the relatively concrete task of improving the machine's search speed and therefore depth, which always improved strength in a predictable way.
How professional chess changed when computers and databases arrived is a useful metaphor for how new technology is adopted across industries and societies in general.
It's a well-established phenomenon. but I feel that the motivations are underanalyzed. Being young and less set in our wavs definitelv makes us more open to trying new things. But simply being older isn't the only factor that works against this openness-there is also being successful, When vou have had success, when the status quo favors vou, it becomes verv hard to volun- tarilv change your ways.
"When will a chess-plaving machine be able to beat the world champion?" This question was put to every chess programmer in his- torv dozens of times. As vou could expect, the earliest predictions, from the days of the digital computer's infancy, were wildly off the mark.
At least the Carnegie Mellon group's daring 1957 promise of 1967 was in some ways avenged since it was a group from the same school whose Deep Blue eventually did the job—if forty vears later instead of ten.
Before that predictable progress phase, there are two previous phases: struggle and then breakthrough. This fits the axiom of Bil Gates, "We alwavs overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten." We expect linear progress, but what we get are vears of setbacks and maturation.
Then the right technologies combine or a critical mass is reached and boom, it takes off vertically for a while, surprising again, until it reaches the mature phase and levels off Our minds see tech progress as a straight diagonal line, but it's usually more of an S-shape.
The limitations of chess weren't the only fundamental misconceptions in the equation. The founding computer science aspect of AI also revealed its limitations. The basic suppositions behind Alan Turing's dreams of artificial intelligence were that the human brain is itself a kind of computer and that the goal was to create a machine that successfully imitated human behavior.
This concept has been dominant for generations of computer scientists. It's a tempting analogy—neurons as switches, cortexes as memory banks, etc. But there is a shortage of biological evidence for this parallel bevond the metaphorical and it is a distraction from what makes human thinking so different from machine thinking
Donald Michie was a machine-learning pioneer himself, applying it to tic-tac-toe in 1960, The basic concept is that you don't give the ma- chine a bunch of rules to follow, the way you might try to learn a second language by memorizing grammar and conjugation rules. Instead of telling it a process, you provide it with lots of examples of that process and let the machine figure out the rules, so to speak.
I'll go into more detail on how computers navigate these openings in the Deep Blue chapters, but I'11 point out now that they rely on a da- tabase of moves derived from human play, called an "opening book." These books have evolved over the years to allow the machines more flexibility, but the basic idea is what it sounds like, a book of openings it follows more or less blindy until it "runs out of book" and has to think foritself, This is effectively similar to how I do it, relying on memory to select the opening lines I prefer until I run out of book and am on my ‣ own,
While the engineering was still a challenge, that dismal equation again illustrates why so many people were disillusioned with how far from its AI roots machine chess had gone. While charting that im- pressive rating climb in 1990, machine intelligence expert and chess master Danny Kopec lamented, "Due to the competitive priorities of most programs, little is revealed about how a program finally selects one move over another, This largely explains why computer chess has appeared to advance primarily as a competitive sport (performance driven) rather than as a science (problem driven)."
That wasn't my last rash piece of rhetoric regarding computer chess, although I would have been fine had I stopped with computers. In an interview around that time, I predicted that a computer would become world champion before a woman did, which turned out to be accurate.
It was interpreted as a sexist slight, which it wasn't. There just weren't any women on the horizon who showed the potential, and that would be the case until the youngest of the three remarkable Polgár sisters from Hungary, Judit, broke into the elite a few years later, eventually reaching top-ten status.