I’ll Die With My Buzzer In My Hand II
Lots in the New York Times yesterday of note on a number of pet projects and peeves, so I’m breaking them up into individual posts. Although “Alex” is on indefinite leave, and along with him my reading on AI, I have to note that IBM’s AI program “Watson,” about which I wrote this post back in April ‘09, is finally ready for his close-up as a Jeopardy! contestant. (I am professionally jealous, having taken the contestant test three times and done well on it, at least on the tangible, intellectual parts if not on whatever intangible qualities they’re looking for in a contestant.) The episodes will air in February pitting supercontestants Brad Rutter and Ken Jennings against the machine. I still predict that Watson will do great on data-crunching questions, such as the geography questions that baffled me last night – which state has the most “O’s” in it’s name? A computer, or a geography nerd, will take about 0 seconds to come up with Colorado, while the rest of us run a finger across our mental map of the US.
But if the writers really want to level the playing field, they will weight more of the questions towards the allusive wordplay for which Jeopardy! is well known. (I can’t count how many times – well, three – I shouted out ‘Egyptian Prescription’ in a contestant search room to the question ‘what Pharoah needs when he has a headache’ or something like that.) Or take the category last night, “Flu Fighters” – a computer program will take that literally, instantly looking in its databases for things about fighting the flu, but a human with adequate cultural references will think of the band Foo Fighters as well – there is no “hard-codeable” connection there, and I can’t imagine the processing power that would support a routine that says something like, “take this phrase and search for everything that sounds the same.” One of the current deficiencies in AI is that it is purely textual – i.e., our brains are “wired for sound,” so we “hear” Foo Fighters when we hear Flu Fighters because our brains are constantly pattern matching sounds to extant data to see if there’s a match. Admittedly software’s intelligence has progressed in the audio department, with the creation of programs that can “hear” a song and tell you the name and artist, but that’s still only data matching, not interpretation. It “knows” Mariah Carey’s “Fantasy” but not that the brilliant riff that carries the song was “sampled” from Tom Tom Club’s “Genius of Love.”
This pattern matching habit works against us sometimes, too. “Closed ears” transform what they hear into what they already know instead of accepting new information. I’ve given up on getting my name right in coffee shops.
“Okay, Orlando, coming right up.” Oh well, it beats “Orly,” or (on Halloween, appropriately) “Merlin.”
So puns, allusions, and other forms of wordplay, it’s safe to predict, will form the bulk of Watson’s “wrong” answers. What remains to be seen, then, is how the writers will, or won’t, tilt the playing field in favor of the humans.