by Edward Goldstick
Please note, dear readers: What follows is a typical "real time" email that I have sent to Jim Fallows on occasion (such as when watching a speech, etc.). Sorry if it seems a bit choppy at times....
OK, it's time for Jeopardy! "The IBM Challenge."
I don't see how I can liveblog on this -- and I don't know how, anyway -- so I'm simply going to take notes and then write it up.
First, the rules as they apply to "Watson" (or I'll call it "W"... or maybe I won't?):
-- Watson cannot hear or see? So, no text recognition? One big strike against Men that I had not anticipated. It will be done via a text file submitted to Watson at precisely the moment that it's shown on the screen. But even so, that's a huge advantage that adds to the latency issue that I previously suggested would be fundamental.
... out of a matrix 6 by 5 = 30 squares ( Man1 / Machine / Man2 )
-- Jeopardy Round: 6 / 17 / 6 (only one not answered)... but it's $2000 / $5000 / $5000
... Sheesh! I keep thinking of Watson in anthropomorphic terms: "he", etc. I'm glad I'm not liveblogging.
... I frankly get the sense that Ken Jennings (Man1) has figured out that he must risk a quicker "button" and then uses an extra moment to think of his response. Am I imagining this?
What? Done already? OK, Double Jeopardy and Final Jeopardy are tomorrow.
And now, a few closing thoughts:
First, last night we watched this very interesting episode of NOVA ScienceNow during which the technology behind Watson was presented in somewhat more explicit terms that those of the show tonight. I recommend it highly -- especially to the degree that it suggests that Watson is really "just" performing a brute force pruning of the decision tree that contains all possible answers.
Second, I'm tempted to tell a story that recounts an experience of an industrial programmer whom I knew in France in the '90s: he had been sent to a 'modern' wholesale fish market in Marseille where the fish lots arrived on a conveyor and the buyers would push a button at their seats when the declining price reached a level that they wanted to accept ... but only one buyer -- the fastest -- would get the deal. My friend was sent because the buyers were complaining that one particular person was getting all the good lots, but he quickly realized that this person always sat in the same seat and that the buttons were polled sequentially with that seat getting polled first after the price changed and they changed after a constant delay which meant that all this buyer needed to do was to push his button in sync with the appearance of the prices as they declined and he would almost certainly win the lot if nobody had purchased it yet. The obvious solution was to randomize the order by which the buttons would be polled.
But the point, if I may, is that it would be really interesting to know whether the buttons on Jeopardy! are polled sequentially (and yes, after all these years, I wonder if they ever checked ... but from my experience, these are the kinds of things that slip through the cracks ... and it may, in fact, not matter significantly). Another question, more interesting in my opinion, would be to know whether Watson is allowed to work on its "pruning" during all down times between questions, not to mention as soon as the topics are exposed at the beginning of each match (and for that matter, it would be really fascinating to know whether the human contestants' minds are "active" both during those interim moments as well as during the evaluation of individual questions ... but its unlikely that the show would be as entertaining if the two human contestants were lying in an fMRI during the competition.)
For what it's worth, I would love to see how these questions would have played out if a fixed delay of 1/2 second were added between the moment that Watson is ready to 'push the button' and the moment that its button is actually recorded as pushed.
Edward Goldstick is a veteran of the high-tech, software, defense, and energy-technology worlds in the U.S. and France.