Winter 2002-2003


 

 

Why My 3-Year-Old Son Is (Far) More Intelligent Than Deep Junior

 

“This is Liav [Goldsmith], and this is Deep Junior.”

Dr. Morris Goldsmith, Senior Lecturer in the Department of Psychology, made a quick comparison of some of the intellectual and cognitive abilities of the two, his young son and a computer developed by two Israelis that is the world computer champ and is intended to play against Gary Kasparov, the world human chess champion:

 

LIAV

JUNIOR

Can perceive (visual, auditory, tactile …)

Can’t

Can use and understand language (speech)

Can’t

Has general reasoning skills

Doesn’t

Has common sense (sometimes)

Doesn’t

Can operate well in a variety of domains

Can’t

Can learn (improve existing skills; acquire new skills)

Can’t

Has self-awareness (incl. metacognition)

Doesn’t

Has other-awareness (incl. theory of mind)

Doesn’t

Understands what he and others are doing, and
(often) why.

Doesn’t

 

Can’t

Can play chess very well

 

    Clearly, as Liav’s father, Goldsmith is more than a little biased, as he readily admits, and perhaps Junior’s “parents” would claim that he has downplayed or omitted some of its abilities – for instance, to calculate the square root of pi.  (On the other hand, Goldsmith told an audience at a symposium on “Man vs. Machine,” sponsored by the University’s Caesarea Rothschild Foundation Institute for Interdisciplinary Applications of Computer Science, “Liav knows what an apple pie is, and Junior doesn’t.”)

    The humor was to emphasize his point.  “It may be possible to pass the Turing test (that is, to achieve intelligent performance that is indistinguishable from human performance) in very specific domains, such as chess,” he said.  “But this is still a far cry from achieving the type of general and specific intelligence that humans possess.  Humans display this intelligence across the entire spectrum of human functioning and in a virtually infinite variety of functional contexts.”

    Goldsmith pointed to one of the most basic features of human intelligence that Artificial Intelligence lacks: understanding.

    He illustrated this with an event taken from a 24-game match played between Deep Junior and its arch computer rival, Deep Fritz, in Cadaques, Spain, in May 2001.  In the endgame, Junior had gained a Rook + Pawn advantage.  There were six pieces left on the board.  “What did Junior do?” the psychologist asked.   “Amazingly, Junior gave away its extra rook to Fritz. Why? Because the resulting 5-piece position was included in its tablebase as a guaranteed win!

    “Technically,” Goldsmith continued, “Junior cannot be faulted with making an error. However, this rather curious move doesn’t leave much doubt that the program plays the game of chess without any understanding of what it is doing.”

    He described another example, this one from game 22. Deep Junior led by two, with just two games remaining. Deep Fritz , showing no understanding of match strategy, decided that this was a good time to offer a draw – which would have virtually clinched the match for Junior.  But, Goldsmith continued, Deep Junior showed even poorer match strategy than its opponent.  It declined the offer—and went on to lose the game and eventually the match.

    This specific omission of match strategy was easily fixed, the psychologist noted, but it was also symptomatic of a much more general, basic deficit.  “Neither Junior nor Fritz have any understanding of the broader context in which they are operating.  They have no idea who or what they are playing against, what their opponent’s strengths and weaknesses are, or how to exploit them.”  No less important for Goldsmith is that they have no idea what their own strengths and weaknesses are.  Therefore, they cannot incorporate such self-knowledge when choosing a line or style of play.

    Goldsmith rhetorically asked whether it was actually necessary to have common sense or any real comprehension of what one is doing and why. He illustrated the general importance of “true understanding” in such an everyday task as taking one’s child to school in the morning.  What if the car won’t start?   GPS (General Problem Solver), a famous AI program of the 1960’s, he related, solved the problem by first taking the car to the garage and having the carburetor fixed and then using “the now smoothly running car” to take the by- now weary, hungry, and very late child to school.

    “GPS simply did not realize that getting the child to school on time, or nearly on time, was more important than getting the car fixed,” he explained.

    Goldsmith summed up his argument on behalf of his three-year-old, “The oversight might be fixed by defining the problem to GPS more carefully. But where would it all end? How many ‘ifs, ands, or buts’ would be needed to take care of every foreseeable and unforeseeable circumstance? And, of course, we’re just talking about one of the many thousands and perhaps millions of tasks that people perform successfully every day: whether it is getting a child to school on time, or getting world chess champion Garry Kasparov to a tournament on time.”

 

 

 

Back to Table of Contents