Email was the user interface for the first AI recommendation engines

Summary

Spinning the radio dials like mini roulette wheels was, in 1993, the best way to discover new music. Static, a snatch of a familiar song, a news report, then ah wait that sounds interesting. You’d walk into a record store with that song stuck in your head, and with any luck would walk out after a conversation with the clerk a few tapes richer, a few dozen dollars poorer.One year later, everything had changed. For in 1994, the best way to discover new music was to email an AI.“This sounds ****in’ moronic,” said one Dave Dell in response to the idea, convinced his love of Lustmord meant the AI couldn’t possibly match his tastes. Even he acquiesced: “I'll try it anyways.”By the time The Cranberries released their hit single “Zombie” that September, over two thousand Daves had tried their luck, emailing Ringo with their favorite artists. To their surprise, they’d get a reply from what appeared to be early artificial intelligence, filled with recommendations of new music they’d love. Enough that, seven years later, science fiction writer Cory Doctorow would reminisce that “half the music in my collection came out of Ringo,” that nascent music AI.Yet incredibly, Ringo was little more than those couple thousand users’ recommendations, averaged and redistributed via email. An email that users would quickly come to think of as a human, a friend, one in a quick succession of email-powered crowdsourced recommendation AIs.The surprising wisdom of the crowdsIt all started with what MIT assistant professor Paul Resnick called “A deceptively simple idea,” in 1994.“People who agreed in the past are likely to agree again,” he postulated, an idea that’d been christened Social Filtering by MIT Thomas Malone seven years earlier. If you and another person both like the same song, or book, or author, there’s a pretty good chance that if one of you likes a new artist, the other will like it as well. The more overlapping agreements you have, the better one’s tastes should be predictive of ...