Robotic domination in I, Robot

Ray Kurzweil's book The Singularity is Near is becoming something of a cult sensation. The 672-page paperback version of the book is ranked 1,494th on Amazon (on par with The Great Gatsby). Recently, Kurzweil announced a Google-backed Singularity University ($25,000 for a 9 week summer program; $12,000 for a 3 day "Executive Program"), lending a touch of academic rigor to an idea that has lived mostly in science fiction. For the time and budget conscious, a rash of Singularity-themed documentaries is now on the horizon.

The Singularity, as I understand it, is the point in time when computers will be smart enough to build even smarter computers, effectively removing humans from the design-build loop of Artificial Intelligence (AI). Kurzweil predicts 2050. That means I'll be 68 when the robots take over!

Predicting the future is no walk in the park, but when it comes to Artificial Intelligence, everyone's packing a lunch. So while I won't try to argue that Kurzweil is wrong (I think he is), it's good to place his predictions in the cultural history of wildly inaccurate AI speculation.

Consider these predictions, both made by outstanding computer scientists actively involved in AI research:

• 1965, Herbert Simon: "machines will be capable, within twenty years, of doing any work a man can do."
• 1970, Marvin Minsky: "In from three to eight years we will have a machine with the general intelligence of an average human being."

As it turned out, these claims were not even remotely true. In fact, the whole history of AI has been one of boom and bust cycles, the product of misplaced exuberant optimism.

Take, for example, the case of machine translation. During the Cold War, the problem of automatically translating intercepted Russian messages received considerable military funding. A 1954 Georgetown-IBM demonstration (translations of 49 chemistry-themed sentences with a 250-word vocabulary) captured public interest and spawned considerable investment, especially as the researchers claimed that the general translation problem would be solved in 3-5 years. When progress turned out to be much slower, funding was cut, and research all but stopped between 1965 and 1993.

Translation research has seen a significant resurgence, especially since I've been in graduate school (for computer science), mostly due to statistical methods. Rather than frame the translation of Russian into English as a series of rules (translate word R3 into word E3; switch the order of words E2 and E4; etc.) written by expert bilingual humans, research consists of building models trained from many examples of translated sentences (word R3 translates to word E3 with probability 0.6; word E3 appears after E2 with probability 0.2; etc.) so that the translation of a Russian sentence is the sequence of English words with the largest total probability, according to the model. The statistical approach is less ambitious-today's models are too simple to capture all of language's nuances-but far more successful.

Kurzweil's Singularity prediction is based on exponential growth. The idea is that because computers have been doubling in speed every two years or so (that's a factor of 1,000 in just 20 years; 1,000,000 in 40 years) huge paradigm shifts are actually quite close. But aside from the issue that computer chips have plateaued due to limits imposed by silicon's insulation ability and the speed of light (new computers have multiple CPUs), progress in automatic translation does not follow the law of exponential progress. Rather, there have been a few periods of dramatic improvement, followed by long periods of very gradual development. This is the trend for the majority of important AI problems.

So, while speculating about the future is both interesting and important, I'd be wary of anyone trying to sell you $12,000 of it.

37.762611 -122.409719

Visualization of possible chess move sequences (try it here)

Artificial Intelligence has always held a special affinity for games. Chess, in particular, was long considered a realm reserved for exquisite human intelligence: the greatest chess players are called Grandmasters; a large percentage of them are eccentric Russian introverts. Gary Kasparov's defeat, by IBM's specialized supercomputer Deep Blue in 1997, was heralded as a major milestone (he contends the match was unfair). But while the dominance of chess-playing software is culturally significant, does it matter for AI?

Chess, like Checkers, Connect-4, and Go, is a game of perfect information. That is, everything useful for choosing your next move is right there on the board (it would be nice to know what your opponent will do next, but you can assume that your opponent is just trying to make the best possible move too). If you had a computer powerful enough, it could consider every possible next move, every possible response, and so on, and finally deduce, absolutely, how to guarantee a particular outcome. To do this is to solve chess, to answer the question: is it possible for white to force a win? Checkers is solved (both players can force a draw). Connect-4 is solved (the first player can force a win). Chess has too many possible board positions to be solved anytime soon.

Deep Blue can compete with human players by searching many moves ahead, testing all possible combinations, and choosing the next move that leaves its opponent with the worst best option. This approach is called minimax search. Since the computer can't search through to all possible checkmates, it searches to a given depth and scores the resulting board position by the pieces each player still has (roughly speaking, a pawn is 1 point, knights and bishops are 3 points each, a rook is 5 points, and the queen is 8 points). Using this rubric, or heuristic, and searching 10-15 moves into the future, makes for an extremely formidable opponent.

Minimax theory was established by John von Neumann in 1928 and the algorithm was improved in the 1950s and 60s to run more efficiently. Deep Blue contains no general innovation that improves significantly on these now classic techniques. The heuristic for evaluating boards has been refined, and the program has a huge database of well-known openings and end-game sequences-when 5 or fewer pieces are left on the board. Thus, Deep Blue is less a marvel of Artificial Intelligence than of engineering: its success is a direct product of the number of positions it can consider in a second (200 million). This is the Brute Force method of problem solving at its finest.

Most real world problems are not like chess. Political maneuvering, for example, is a game of imperfect information, where each player must guess at underlying motives and resources from superficial clues. The language of political, and in particular war-time gamesmanship, has shifted markedly away from chess… towards poker. Obama tipped his hand, Chavez is bluffing, Ahmedinejad is all in.

And Artificial Intelligence for poker is still far behind humans. The University of Alberta's Polaris system earned a narrow victory at the 2^nd man-machine poker match last July, but the competition involved heads-up limit poker: one-on-one games where the only possible bets are $10 or $20. Compared with the main event at the World Series of Poker, which has no betting limit, and about 10 players at one table, this is something of a "toy" problem. Recent research focuses on how to model opponents-that is, automatically refining the software's understanding of the meaning of each players' bets as information is gathered about how those players play.

Over the next decade, I would guess that poker research, perhaps backed by military funding, will expand significantly. And unlike Deep Blue, poker software that can dominate a table full of professional players, will be the product of significant advances in the field of Artificial Intelligence.

37.762611 -122.409719

Bio-inspired design borrows its creative inspiration from models and systems in nature, that is, plant and animal parts that have been slowly tweaked for over 3.8 billion years. But that doesn't mean that nature's designs are perfect. In fact, that's what makes the process of engineering things based on natural models so difficult. You have to figure out how to pull the aces from the evolutionary discard pile. As professor Bob Full at U.C. Berkeley explained in our first phone conversation, that's also why scientists now use the term "bio-inspiration" rather than the more commonly known term "biomimicry." Biologists and engineers are not looking to simply mimic nature, because there are all kinds of dead ends and redundancies in natural systems that would be pointless to recreate in an optimized, man-made piece of technology. One of the examples he gave me is a kind of grasshopper that if you were to copy it, you would copy neurons that go to nothing, they don't connect to any muscles, and that's because during evolution the adults lost their ability to fly. The neurons going to the muscles are still there, but the muscles aren't there anymore. No need to copy that, right?

So what a biomimeticist does is look to nature to find plants & animals with remarkable performance abilities, and studies their adaptations for inspiration to design something new. For example, if you want to make a tiny robot that can fly, then look at the best fliers. If you want to design a blade that moves quickly through fluids, or an Olympic swimsuit that minimizes drag, then look to the most efficient swimmers. Now that's what I call "intelligent design!"

Watch the Bio-Inspiration: Nature as Muse television story report online.

37.871754 -122.260760

Though it was McCarthy who persuaded his nine other colleagues at the conference to adopt the term "artificial intelligence" to describe the nascent field, the seeds of artificial intelligence were planted earlier. Alan Turing, who was instrumental in breaking the German's Enigma code during WWII, published a paper in 1950 that laid out what came to be known as the "Turing Test:" if a machine could carry out a conversation with a human in such a sophisticated manner as to trick the human into thinking that he or she was conversing with another human, then the machine would have displayed true "intelligence."

But nearly 60 years later, the world still awaits a machine capable of exhibiting "general A.I.", instead of the "narrow A.I." demonstrated by IBM's chess-playing Deep Blue or Stanford University's Stanley, an autonomous robotic vehicle, or other impressive albeit limited applications of A.I. For example, Deep Blue may be able to beat Gary Kasparov at chess but can it beat a 10 year-old at a game of checkers? The lack of a general A.I. is made even more stark when juxtaposed with Moore's Law, a maxim that goes back to 1965 when Intel founder Gordon Moore postulated that the number of transistors on a computer chip would double roughly every 18 months.

There's even a term – "Singularity" – that is being used to describe the moment when technological progress will leapfrog and herald the creation of computers that not only achieve human-like intelligence, but also give rise to a progeny of computers who will be smarter then their digital forbears. Though he didn't coin the term (sci-fi writer Vernor Vinge did), the most famous exponent of this belief is inventor Ray Kurzweil. He places the Singularity as occurring sometime before 2050 and believes that with the advent of this unheralded technological progress, mankind may solve some of our society's most pressing ills, such as global warming, and even conquer death, by uploading one's consciousness into a virtual medium.

Though this seems a far stretch from engineering a domestic robot like Stanford's Artificial Intelligence Robot, top A.I. researchers like Stanford's Andrew Ng and Daphne Koller do believe that computing systems will some day be as smart or smarter than humans. When I spoke with Dharmendra Modha about his work into cognitive computing at IBM, he talked effusively about creating an "i-Brain," a digital accessory that people could carry around, making decisions and processing information like its human cousin. But if you're like me, and lament those moments when you've misplaced your keys or other instances of poor neural performance, you can't help but think that such a device can't arrive soon enough. On second thought, I'll wait until v2.0 hits the shelves.

Watch the Artificial Intelligence: Thinking Big television story report online.

And don't miss our Web Extra: A Dose of A.I. In this QUEST web exclusive, Stanford University computer science professor and artificial intelligence (A.I.) researcher Daphne Koller provides an elegant explanation of how A.I. can be employed in the examining room to diagnose a patient's illness more accurately than a human clinician. Find out more and learn how medical diagnosis is just the tip of the iceberg when it comes to tasks that rely on making sense of a sea of data to arrive at an informed conclusion.