"Watson" to play Jeopardy 2/14-2/16

[Mrbloodworth]

Hes a contained system. No web crawling.

[Merusk]

(I for one welcome our new computer overlords)

Jennings actually wrote that as part of Final Jeopardy last night.  Was pretty funny.

Is Watson really "smarter" or just quicker with the buzzer?

For certain definitions of smarter, it's smarter.  (In the same way those high-functioning Autistics who can spit out huge math equations are smarter)   In the end it doesn't matter, because computers WILL wind-up smarter than us within our lifetimes.   The edge we have (as has been pointed out in a few articles this week) is creativity.  We can't figure out how to make machines more than processors.  Yet.

[Hawkbit]

Hes a contained system. No web crawling.

I only watched parts of it - seriously?  They didn't pull from the web at all?  That just seems... dumb.  They basically programmed a talking encyclopedia that gets shit wrong then. 

[Mrbloodworth]

Hes a contained system. No web crawling.

I only watched parts of it - seriously?  They didn't pull from the web at all?  That just seems... dumb.  They basically programmed a talking encyclopedia that gets shit wrong then. 

[Merusk]

Hey, if it doesn't work 100% perfectly on the first time out it's clearly a failure and waste of resources.   

[Mrbloodworth]

They likely did it a favor by not exposing it to the web. Can you imagine? They let anyone and anything on the web. Every third answer would have been shitcock.

[bhodi]

I only watched parts of it - seriously?  They didn't pull from the web at all?  That just seems... dumb.  They basically programmed a talking encyclopedia that gets shit wrong then.  

That is a gross misstatement of the technical subtleties involved. It's a self-contained mirror of large, pertinent parts of the internet, pre-selected based on the jeopardy category and indexed similarly to the way google crawls the real internet. Don't discount the amount of technical sophistication, because developing algorithms and the backend to provide near realtime indexing made google billions upon billions upon billions of dollars.

It isn't just a talking, fallible encyclopedia. It's really more like an entire search engine server farm, only smaller, because the subset of pertinent data is less and it only has to parse one query at a time.

[Trippy]

There were some questions on the second night where Watson had a fairly positive idea of the answer, but didn't buzz in first. It appeared that questions that didn't have straightforward and obvious keywords took Watson a little longer to sort through. I missed the first night, so I didn't get all the details. Is it running off of an inhouse db, or is it actually webcrawling?

And yes, I know succeeded was wrong (who would ever intentionally use that word in reference to Toronto?)

Watson wasn't connected to the Internet. It had its own database.

[pxib]

"Watson has lots in common with a strong human 'Jeopardy!' player: it's very smart, very fast, speaks in an uneven monotone, and has never known the touch of a woman."

[Jeff Kelly]

The human mind works context based and associative.

The big step forward for me lies in the natuarl language processing capabilities of Watson.

Human Language is very context sensitive. I'm not even certain that a grammar for any human language would even lead do decidable/computable problems.

Watson has to actually parse a spoken question decide on the most likely context that question is phrased for and then use some for of associative memory to infer the answer.

If you would ask the question to a human operator that knew the right context and were able to then search for the question on google the most work would have already been done once the operator enters the search terms.

So I'm less impressed by the data mining capabilities of Watson (that could be brute forced if necessary) but by actually arriving at the right terms - given a deliberately obscure question - as input for the search given only a spoken query in human language form.

Parsing human language and context selection is one of the hardest comp-sci problems around (PSPACE even, IIRC) and many pure computing approaches are unfeasible because it either takes to long to arrive at the 'correct' context and sentence or the grammar is even undecidable.

So that Watson is able to understand human language well enough to actually play Jeopardy! is an achievement in its own right. The associative memory and efficient lookup is just the icing on the cake.

[LK]

"Watson has lots in common with a strong human 'Jeopardy!' player: it's very smart, very fast, speaks in an uneven monotone, and has never known the touch of a woman."

[Teleku]

Ok, that actually made me lol here at work.  Now there on to me that I might not actually be working on the web site.  Thanks!

[Ingmar]

I only watched parts of it - seriously?  They didn't pull from the web at all?  That just seems... dumb.  They basically programmed a talking encyclopedia that gets shit wrong then.  

That is a gross misstatement of the technical subtleties involved. It's a self-contained mirror of large, pertinent parts of the internet, pre-selected based on the jeopardy category and indexed similarly to the way google crawls the real internet. Don't discount the amount of technical sophistication, because developing algorithms and the backend to provide near realtime indexing made google billions upon billions upon billions of dollars.

It isn't just a talking, fallible encyclopedia. It's really more like an entire search engine server farm, only smaller, because the subset of pertinent data is less and it only has to parse one query at a time.

Plus, and this is very important, it doesn't just give one answer really. It was giving just one on Jeopardy because that's what the format requires, but it can output a list with weighted probabilities, which in a lot of ways is much more useful. There's a lot of interesting applications for this.

[Lantyssa]

Well, if it was getting the answer correct, it was still impressive.  Think about Google's "I feel lucky" button giving back a good answer to a every natural language query.