In The Armchair

Watson and the Singularity – IV

Posted in Computers by Armchair Guy on February 17, 2011

On the last day, Watson was not nearly as dominant as on the second day, but won a big packet on the Final Jeopardy question to take its total winnings to much higher levels than either of its contestants.  For a large part of the third day, Ken Jennings was in the lead.

Watson uses the correctness of previous answers to try to understand what a category means.  For example, in a “Name the Decade” category, Watson wasn’t sure what the category name meant — so it didn’t know that it could restrict its answers to the set {1900, 1910, 1920, …, 1990, 2000}.  When it is unsure in such a manner, Watson uses previous correct answers in that category — possibly by its competitors — to narrow down what the category means.  It was clear on the third day that Watson didn’t understand what certain categories meant, even after observing opponents’ correct responses.  This meant it did poorly on those categories as a whole.

Watson’s natural language processing, I think, is tailored to the task at hand — winning Jeopardy.  Like I said before, it doesn’t understand sentences the same way a human would.  While watching the show, I found that many of the answers were found in the intersections of two or more sets, but Watson didn’t identify all of the sets.

I would classify Watson as a mild sub-Singularity event at this point.  If indeed programs like Watson proliferate the way chess programs have — if Watson clones become much more powerful and lightweight enough to run on the computer as personal assistants, perhaps with some help from the cloud — we will be on our way to real artificial intelligence.  Sequential improvements in such programs will eventually lead to super-human intelligence, much as the Singularity gurus predicted.  Eventually, APIs for NLP and this type of reasoning might become commoditized — unless companies like Google prefer to provide access APIs only, and keep all the computation hidden on their servers.

The question that immediately popped into my head when I heard about this Jeopardy challenge for the first time was “why not Google?”.  IBM has a fantastic record of innovation, of course, but the things Watson does are right up Google’s alley.  Search would be greatly improved if you could ask a question and have it answered in addition to being served a bunch of related pages.  Personally I believe Google has already developed a system like Watson.  So why don’t we know about it?  Two possibilities.

First, Watson needed 2800 processors to answer questions one at a time.  The technology that Google has may or may not be equally advanced, but perhaps doesn’t scale up to allow answers for millions of questions yet.

Second, this is a card in Google’s hand that it doesn’t want to show unless necessary.  If a competitor (mostly Bing at this stage) appears to be making significant inroads into its search space, it can add this feature to jump ahead, so it’s insurance.  Revealing everything would just provide Microsoft with a “copy this!” blueprint.