Will Watson win on Jeopardy? -- 02/09/11
You may have already seen promos or news stories about the IBM computer nicknamed "Watson" that is challenging the two top Jeopardy champions to a match, but as the broadcast comes closer I just can't resist commenting on it and on IBM's coming 100th birthday celebration.
In 1997 an IBM computer named Deep Blue defeated Gary Kasparov, the reigning human chess champion, in a six game match (two wins for Deep Blue, one for Kasparov, three draws).
Now IBM has come up with a new machine -- named Watson in honor of Thomas J. Watson, IBM's founding CEO -- designed to play Jeopardy. It is a huge challenge for a computer because the game of Jeopardy works with an unlimited range of subject matter and employs clues that include word games, puns, riddles, irony, jokes, and subtle shades of meaning -- all things at which humans do well and computers do not (as can be attested by anyone who has been reduced to cursing at an uncooperative PC). This will be an opportunity for IBM to show off tremendous advances in systems design, deep analytics capability, and natural language parsing and processing... or an opportunity for deep blue embarrassment on national television.
For those of you in the U.S., Nova (on PBS tonight) will devote an entire program to Watson, calling it "The Smartest Machine on Earth." (10 pm EST, the schedule in your area may vary.) The special Jeopardy contest between Watson and the two all-time human champions (Ken Jennings, who won a record 74 consecutive games, and Brad Rutter, who won more than three and a quarter million dollars) will be broadcast on February 14, 15, and 16. (That would be 7:30 pm on channel 12 in Providence; you'll have to check your local listings.)
Oh, about that 100th birthday thing. In June of 1911 in Endicott, NY, a new company was formed: the Computing, Tabulating, Recording Corporation (via a merger of the Computer Scale Company of America, the International Time Recording Company, and the Tabulating Machine Company). This was the birth of IBM, although years would go by before CTR changed its name to International Business Machines.
IBM commissioned a Centennial Celebration film -- 100 x 100 -- and that is a link to YouTube where you can view it. It's thirteen minutes long but it is delightful, featuring 100 people, beginning with a man who was born the same year as IBM. Look at this this way: IBM has the money to hire the creative people to make an interesting film. If you've got a broadband connection, it is worth checking it out.
Twenty years ago IBM seemed to be in trouble, to have lost its corporate way, but then Lou Gerstner came in as CEO and turned things around (the title of his memoir of his years at IBM is Who Says Elephants Can't Dance) and IBM has continued to be quite successful under current CEO Sam Palmisano.
A few months ago there was a news story out about China producing a new supercomputer -- the Tianhe-1A -- that was the fastest in the world at 2.57 petaflops. IBM is now building Mira, a supercomputer for the U.S. Dept. of Energy's Argone Laboratory, that when it is completed next year will run at 10 petaflops, almost four times faster than the Chinese machine.
What is a petaflop? Well, one way of measuring computer performance is in terms of how many floating point math operations it can execute in one second -- that measure is called a flop. (What is floating point math? Remember what they probably called "scientific notation" when you were in school? This is sort of a way of doing that in a computer - it allows for a wider range of number values to be represented more efficiently.) 1 million floating point operations per second = 1 megaflop; one thousand times that, 1 billion operations per second = 1 gigaflop; one thousand times that, 1 trillion operations per second = 1 teraflop; and one thousand times that, 1 quadrillion operations per second = 1 petaflop.
Back in my computer operator days, the university where I worked had an IBM 370/158 mainframe computer that was rated just under a megaflop -- one million floating point operations per second -- if I recall correctly it was somewhere around .8 megaflop -- so 800,000 of those floating point calculations per second). In other words, Mira will be more than twelve and a half billion times faster than that old mainframe. If you had every single man, woman, and child in the United States perform one floating point calculation per second -- every second of every minute of every hour of every day non-stop for an entire year -- they would perform the work that Mira will do in one second.