HOUSE_OVERSIGHT_018419.jpg

Chapter Eleven:
Citizens!

In which the Seventh Sense rescues us from an unexpected danger.

1.

I never needed much incentive to go see Pattie Maes. Belgian, usually dressed in
some black fashionable getup, she was like a human shot of espresso. You ended
every conversation wide awake, eyes open. When I first met her in the 1990s, she
was in charge of much of the work on artificial intelligence at MIT’s Media Lab –
Danny Hillis’ old home. Maes had arrived at MIT in 1993 and almost immediately
turned to the problem of making machines that might think. One day, as we were
discussing just how the strange miracle of computer thought might occur, she
introduced me to a puzzle of her field that has stayed on my mind in the years since.
It is called the “Disappearing AI Problem.”

Back in the 1990s, as the Internet was emerging into popular consciousness, Maes
and her team were tinkering with what was known as computer-aided prediction.
This was an advance on the ping-pong conversations Joseph Weizenbaum had
coerced from ELIZA in the 1960s, You: “I am bored.” ELIZA: “Why are you bored?” In
Maes’ experiments a computer would ask, for instance, what movie stars you liked.
“Robert Redford,” you’d type. And then the box would spit back some films you
might enjoy. Cool Hand Luke. And, well, you had liked that film. This seemed like
magic at the time, just the sort of data-meets-human question that showcased a
machine learning and thinking. An honestly “artificial” intelligence. Maes hoped to
design a computer that could predict what movies or music or books you or I might
enjoy. (And, of course, buy.) A recommendation engine. We all know how sputtering
our own suggestion motors can be. Think of that primitive analog exchange known
as the “First Date”: Oh, you like Radiohead? Do you know SigurRos? Pause. Hate them.
Can you really predict what albums or novels even your closest friend will enjoy?
You might offer an occasional lucky suggestion. But to confidently bridge your
knowledge of a friend’s taste and the nearly endless library of movies and songs and
books? Beyond human capacity. It seemed an ideal job for a thoughtful machine.

The traditional approach to such a problem was to devise a formula that would
mimic your friend. What are their hobbies? What areas interest them? What cheers
them up? Then you’d program a machine to jump just as deep into movies and
music and books, to break them down by plot and type of character to see what
might fit your friend’s interests. But after years building programs that tried – and
failed – to tackle the recommendation problem in this fashion, the MIT group
changed tack. Instead of teaching a machine to understand you (or Tolstoy), they
simply began compiling data about what movies and music and books people liked.
Then they looked for patterns. People were not, they discovered, all that unique.
Pretty much everyone who liked Redford in Downhill Racer loved Newman in The
Hustler. Anyone who enjoyed Kid A could be directed safely to (). Maes and her team
found themselves, as a result, less focused on the mechanics of making a machine

187

HOUSE_OVERSIGHT_018419