on my book about IBM's Jeopardy computer, the most common
question I encounter is this: Doesn't Google already answer questions?
The short answer is no. Google depends on our brains in two ways: It
gets us to think like a computer when formulating our query, picking
three or four words that will make most sense to the machine. Then it
directs us to the neighborhood of the answer we're looking for, but
leaves it to our infinitely more nuanced brains to find exactly what
we're looking for there.
Watson, which will face off
against two Jeopardy legends, Ken Jennings and Brad Rutter, in
February, has to handle all that work by itself. It must decipher
complex English, hunt down possible answers, choose one, and decide if
it has enough confidence to bet on it.
Here's an example: "When 60 Minutes premiered, this man was
U.S. president. That's a tough one for a computer. It has to understand
what "premiered" means and that it's associated with a date. Then it has
to figure out the date when an entity called 60 Minutes
premiered, and then find out who was U.S. president at that time. In
short, it requires a ton of contextual understanding -- or a statistical
simulation of it -- and then two different hunts, one for the date, the
second for the president.
Once Watson has a list of possible answers (or "responses," as they
call them on Jeopardy!), it has to figure out which one merits the most
confidence, and if it's sure enough of the answer to place a bet on it.
All these takes place in about 3 seconds. (By the way, the answer is
Watson has more than 100 algorithms leading it to solve these
Jeopardy clues. Each one has its specialty. One of them helps it with
specifically this kind of question. It's called "nested decomposition,"
and it involves breaking the clue into two different hunts. This may
sound really obscure, and only marginally useful. But if you listen to
people asking questions, a lot of them require this type of hunt.
"What's the best pizza joint near campus? Which southern state has a big
steel industry? etc.
So you might think that Watson could become the next Google. One big
problem. To solve that clue, Watson uses more than 2,000 processors and
consumes loads of electricity. Google, in those same three seconds,
responds to millions of search queries. Google uses perhaps
one-billionth of Watson's resources, or less, to handle each query. So
the two approaches don't compete. But in coming years, Google and the
other search engines will start to answer questions, more like Watson.
(They're starting with simple ones. Nested decomposition is still a ways
off.) And to get Watson jobs outside of show biz, IBM will have to
figure out how to run such machines for a fraction of the cost.
(first published on The Huffington Post)