Speed is not an issue here in any way. If you can get a computer to perform a task in 20 seconds, then you can also get the computer to perform the same task in a fifth of a second, just by the trivially simple expedient of making the computer 100 times faster. And this is one of the world’s largest computer companies we’re talking about, here: They’ll be stuffing this thing full of all of the fastest hardware available anywhere, and damn the cost.
Simplicio, that MIT language bot is hilarious-- I’ve gotten better answers from a Magic Eight Bob. I tried to throw it a bone by asking “How big is an acre?”, but it couldn’t even handle that.
Oh, spiffing. I finally got an audition, not even a week ago. They told us to smile, to show lots of personality. Now they’re going to let a freakin’ machine on?
One of my all-time favorite Jeopardy moments was an audio Daily Double in “Before and After”. I don’t recall what the spoken clue was (it was a hint to the name of the band), and then they played a snippet of Take the Skinheads Bowling.
I’ll be curious to see how the computer handles the audio and video clues. And if they have to eliminate those to accomodate the computer, that would be an act of craven puppyhood.
For that matter, I wonder how the computer decides which category to select.
Which is why data organization would be a big part of any successful system. It’s been a while since I looked at AI, but organizing data using knowledge of the world, something we do very well, would be a major advance - especially since they’d have to find some paradigms for collecting and arranging data. I doubt IBM has enough invested in people to make manual sorting of poems by the fame of the poet (and by type, period, and a zillion other relationships I can think of) practical.
Yea, manual sorting is certainly out. But you can simply take a bunch of traffic data. For example, Wikipedia has a poets category, and I bet the 100 most visited pages in that category are close to the 100 most popular poets. Similarly, famous poets probably have more links to their article from other articles then non-famous ones.
It would basically be a variation of Google page rank. If two articles gave different answers about a question, the program would choose the one that came from the most popular page since Trebek is more likely to ask a question about something famous then something obscure.
I asked it “Can entropy ever be reversed?” and it said “I don’t know that.” Lame. Someone should have told it to answer “THERE IS INSUFFICIENT DATA FOR A MEANINGFUL ANSWER.” I’m formally disappointed with MIT programmers now.
Sorry to drag this thing back up (and possibly inadvertently ringing in the zombie sentient computer apocalypse), but Watson has just bested two human opponents – Brad Rutter and Ken Jennings – in a first training match. The official match (which is being recorded today) will be broadcast a month from now, on Feb. 14. Story here.
Personally, I’m pretty impressed by IBM’s accomplishment – understanding the subtleties, inaccuracies and often context-dependent meaning of natural language is notoriously hard to do for a computer, and mastering the often pun-laden statements in Jeopardy is a top-level challenge in that regard.
As a bonus nugget, when asked about Watson’s state of mind prior to the big match, its lead designer, David Ferrucci, said:
When I first saw this headline online, I thought, @#)(*$#)(ing spoilers, only to be relieved it was only a test game. I hope the results of the real match are successfully unspoiled until next month!
I’m just so, so disappointed Watson wasn’t named, oh, Maximillian. Say, maybe the fifth iteration of that model. That test run could have been much more entertaining.
