[ExI] Watson On Jeopardy.

[Richard Loosemore]

David Lubkin wrote:
> Richard Loosemore wrote:
> 
>> AI researchers have, over the years, publicized many supposedly great 
>> advances, or big new systems that were supposed to be harbingers of 
>> real AI, just around the corner.  People were very excited about 
>> SHRDLU.  The Japanese went wild over Prolog.  Then there was the 
>> "knowledge based systems" approach, aka "expert systems".  Earlier on 
>> there was a 1960s craze for "machine translation".  In the late 1980s 
>> there were "neural networks" vendors springing up all over the place.  
>> And these were just the paradigms or general clusters of ideas ... 
>> never mind the specific systems or programs themselves.
>>
>> Now, the pattern is that all these ideas were good at bringing down 
>> some long-hanging fruit, and every time the proponents would say "Of 
>> course, this is just meant to be a demonstration of the potential of 
>> this new technique/approach/program:  what we want to do next is 
>> expand on this breakthrough and find ways to apply it to more 
>> significant problems". But in each case it turned out that extending 
>> it beyond the toy cases was fiendishly hard, and eventually the effort 
>> was abandoneed when the next bandwagon came along.
> 
> I think it was my graduate advisor, Carl Page, who told me that whenever 
> you saw a paper in AI with "examples" of sentences understood, theorems 
> proven, etc., the "examples" were in fact the *only* ones the system 
> could cope with.
> 
> (I was predisposed to be enraptured by AI, but his doctoral classes were 
> the best I've taken in any field.

Indeed.

What puzzles me about this whole Watson discussion is that the skeptical 
perspective I have been presenting here is, well, something that many 
people in the cognitive science world of about 10 to 20 years ago would 
have considered a no-brainer.  And many in AI, too, paradoxical as that 
might seem.

*Many* people working in those fields, back in the late 80s, were 
becoming frustrated with AI systems and AI claims that were overblown. 
Just about everyone was talking about the silliness of programs that 
could handle a selected examples, but which were not generalizable.

The basic mechanism under the Watson hood was known quite some time ago, 
and it was kind of obvious that if you went to the trouble of stuffing 
it with a huge amount of data, you could do something superficially 
impressive like handling Jeopardy questions.  But why bother to do that 
when it did not address the underlying issues?  Issues that were known 
decades ago.  Beats me.

Anyhow, seems I have to take the rap for saying what was common knowledge.

Tell you what, hang on folks and I will do a quick survey of my 
cognitive science friends to see what they think about Watson as an 
advance in AI. Get back to you on that.


> My parenting philosophy began with a conversation with him. Dr. Page 
> sent his two sons to a Montessori school. He extolled the value of 
> Montessori in raising a strong, confident child who thinks for himself, 
> but warned me that it comes at a price: When you have a Montessori kid, 
> they will never accept "Because I say so."
> 
> The consequences of his parenting style are that his namesake, Carl Jr., 
> co-founded eGroups, which was sold to Yahoo for 0.5 billion and we now 
> know as Yahoo Groups. The other son, Larry, co-founded Google, and is 
> worth $15 billion.)

Great story!

You have suddenly made my frustratingly independent-minded son, who 
argues about everything under the sun, and who we sent to Montessori, a 
lot easier to deal with.... ;-)


Richard Loosemore