Re:IBM Watson: Overprovisioned “Big Iron”?
by burke on Wednesday February 16, 2011 @11:08PM
(#2229701)
First off, I think it's naive to assume it would be that easy to do as well as Watson. Talk is cheap.
More importantly, though, Watson isn't /really/ about solving Jeopardy. It's a more general AI designed to provide concise answers to arbitrary questions posed in arbitrary natural language. I have absolutely no source for this, but it's probably reasonable to assume that Watson has the most sophisticated understanding of human language of any computer system to date. This is much bigger than a quiz show.
Re:IBM Watson: Overprovisioned “Big Iron”?
by gojomo on Thursday February 17, 2011 @12:18AM
(#2229881)
Indeed, talk is cheap. But no other team was even given a crack at this, nor have I heard of any research groups trying (and failing) at similar trivia tasks. Maybe we'll see that now.
Re:IBM Watson: Overprovisioned “Big Iron”?
by phaedrus on Thursday February 17, 2011 @02:13AM
(#2230173)
Then by all means build such a single-server Watson clone and program it!
I think you're underestimating the combinatorial explosion that occurs when you try to do any kind of AI over natural language (and I speak from experience). Though I get your point that it seems they haven't reduced the big-O complexity of the problem; but what if it's irreducible?
In fact I do think eventually a smaller scale implementation of a Watson clone will happen, and not just because of Moore's law. As Steve Wozniak said in a forward to a book about the Apple 1, "Some very simple ideas are very hard to do the first time." It seems to take a first execution to inspire the people who will come after and find simpler ways to do it. Somehow just knowing that it has been done is enough to set another inventor down the path of doing it even better. This happened frequently in the field of steam engines and thermodynamics in the 17 and 1800s, when just hearing about what one person had done would lead the next person to leap frog it.
Re:IBM Watson: Overprovisioned “Big Iron”?
by spitfire on Thursday February 17, 2011 @02:25AM
(#2230189)
Michael Abrash said something like Woz's quote also. Once you know it /can/ be done it somehow becomes easier. Often the less you know the better you are.
Example, in his black book abrash tells a story about working at number nine. Their got word that one of their competitors had put a FIFO buffer on one of their cards (this was the ISA period 386 era). The staff at number nine panicked. They were so close to being finished their design they didn't have the space for a FIFO. So they put a single bit write FIFO in.
Then they taped out and went to market, praying they could compete. Then the other card came out on the market.... with an 8byte read FIFO. At the time video memory was /dead/ slow, so everyone designed their programs to never, ever, ever read VGA memory. Instead you wrote to a virtual buffer and flipped it to the screen.
The results, number nine absolutely creamed the competition for some time. And people still didn't read VGA memory.
Moral: sometimes just knowing it can be done is all you need. and often too much knowledge will send you down the wrong path.
Re:IBM Watson: Overprovisioned “Big Iron”?
by gojomo on Thursday February 17, 2011 @02:51AM
(#2230236)
Then by all means build such a single-server Watson clone and program it!
What's the prize? Can I get the same access IBM had, including a show appearance if my implementation is competitive enough?
(I'm not saying the task is easy enough for an unpaid hobbyist to attempt; I am saying IBM's 10 racks of 9 maxed-configuration machines seems more designed for promotional optics than to match the task at hand.)
Re:IBM Watson: Overprovisioned “Big Iron”?
by spitfire on Thursday February 17, 2011 @12:00AM
(#2229838)
Well, Watson was written in Java....
Seriously though, While you could write a simple wiki search engine to get decent results right away I think IBM is playing the long game. Taking the much more difficult (15 people 5 years) game for long term results. The begining results like this jeopardy game don't seem that impressive, but if the watson technology is as impressive as ibm make out I'm sure it'll broaden in time.
On the other hand, I found the use of power chips over x86-64 a bit suspect. That was clearly IBM sales at work.
Re:IBM Watson: Overprovisioned “Big Iron”?
by gojomo on Thursday February 17, 2011 @12:30AM
(#2229909)
If another IBM engineer walked in on the Watson team and showed them his single-server (32 core, 512GB RAM) alternative that performed on par with the Big Iron approach, do you think IBM corporate's reaction would have been:
• "Great, now we can haul the unit to Jeopardy's Burbank studios for a showdown ASAP!" ...or...
• "Hold that thought for the future, we've already frozen the datasheets and shot lots of footage of the server room that's perfect for the show and B-rolls."
Re:IBM Watson: Overprovisioned “Big Iron”?
by spitfire on Thursday February 17, 2011 @02:18AM
(#2230181)
I agree with you. Don't change horses midstream. I think the choice to go with power was flawed to begin with. But for better (financially) or worse they decided to do that.
Mind you, several racks worth of computer always looks cooler. You'd be a bit disappointed if Alex Trebek opened the rack door to find a 1u Dell inside.
Re:IBM Watson: Overprovisioned “Big Iron”?
by icandoitbetter on Thursday February 17, 2011 @01:19AM
(#2230020)
Well, he did search Google with the Jeopardy questions and showed that very often the answer to the question was Google's first result. I think that's not only indicative of the difficulty of the problem at hand, but also shows that Watson is no different than established A. I. systems such as search engines. I had never seen Jeopardy before (not being from the States) and expected the game to be slightly more challenging; too many clues are included to the question asked and they make it too easy to determine the type of the expected answer ("this" appears almost in every single question). All three players seem to know the answer to any given question, and it's just a game of who presses the button first.
More importantly, though, Watson isn't /really/ about solving Jeopardy. It's a more general AI designed to provide concise answers to arbitrary questions posed in arbitrary natural language. I have absolutely no source for this, but it's probably reasonable to assume that Watson has the most sophisticated understanding of human language of any computer system to date. This is much bigger than a quiz show.
I think you're underestimating the combinatorial explosion that occurs when you try to do any kind of AI over natural language (and I speak from experience). Though I get your point that it seems they haven't reduced the big-O complexity of the problem; but what if it's irreducible?
In fact I do think eventually a smaller scale implementation of a Watson clone will happen, and not just because of Moore's law. As Steve Wozniak said in a forward to a book about the Apple 1, "Some very simple ideas are very hard to do the first time." It seems to take a first execution to inspire the people who will come after and find simpler ways to do it. Somehow just knowing that it has been done is enough to set another inventor down the path of doing it even better. This happened frequently in the field of steam engines and thermodynamics in the 17 and 1800s, when just hearing about what one person had done would lead the next person to leap frog it.
Example, in his black book abrash tells a story about working at number nine. Their got word that one of their competitors had put a FIFO buffer on one of their cards (this was the ISA period 386 era). The staff at number nine panicked. They were so close to being finished their design they didn't have the space for a FIFO. So they put a single bit write FIFO in.
Then they taped out and went to market, praying they could compete. Then the other card came out on the market.... with an 8byte read FIFO. At the time video memory was /dead/ slow, so everyone designed their programs to never, ever, ever read VGA memory. Instead you wrote to a virtual buffer and flipped it to the screen.
The results, number nine absolutely creamed the competition for some time. And people still didn't read VGA memory.
Moral: sometimes just knowing it can be done is all you need. and often too much knowledge will send you down the wrong path.
What's the prize? Can I get the same access IBM had, including a show appearance if my implementation is competitive enough?
(I'm not saying the task is easy enough for an unpaid hobbyist to attempt; I am saying IBM's 10 racks of 9 maxed-configuration machines seems more designed for promotional optics than to match the task at hand.)
Seriously though, While you could write a simple wiki search engine to get decent results right away I think IBM is playing the long game. Taking the much more difficult (15 people 5 years) game for long term results. The begining results like this jeopardy game don't seem that impressive, but if the watson technology is as impressive as ibm make out I'm sure it'll broaden in time.
On the other hand, I found the use of power chips over x86-64 a bit suspect. That was clearly IBM sales at work.
• "Great, now we can haul the unit to Jeopardy's Burbank studios for a showdown ASAP!" ...or...
• "Hold that thought for the future, we've already frozen the datasheets and shot lots of footage of the server room that's perfect for the show and B-rolls."
Mind you, several racks worth of computer always looks cooler. You'd be a bit disappointed if Alex Trebek opened the rack door to find a 1u Dell inside.
How does it show that?