Samantha: “And then I had this terrible thought. Are these feelings even real? Or are they just programming?”
An anodyne and restrained Theodore Twombly falls in love with Samantha, the female persona of his computer operating system, in the recent movie Her. The romance begins as they overcome their somewhat awkward initial encounter and settle into an easygoing relationship. She arranges his life and tries to fix him up with a date. He tells her about his dreams. They banter, bond over his ongoing divorce and have endless conversations about people, events and desires as they explore the Los Angeles of the near future. Little by little, she reels him in.
Theodore's face lights up every time she calls. He is clearly in love with this ethereal being. He panics when he can't talk to her. The honeymoon ends after a disastrous attempt at a ménage à trois. Their deteriorating relationship is finally rent asunder after she confesses that she is carrying on intimate conversations with hundreds of others.
What is remarkable about this smart movie is that the story is so convincing. It leads you to believe that a man can fall in love with a disembodied female voice, an applet. While Theodore is flesh and blood, Samantha is a software program. She is linked to him by his omnipresent phone, a kind of glorified Siri living in the cloud, where she is simultaneously interacting with thousands of other people.
Her does not delve into the profound questions that the existence of a Samantha implies: Can she really love Theodore, or is she just feigning it? Does it feel like anything to be her? Indeed, can a software construct ever be conscious, or is it condemned to a zombie existence, cleverly programmed to respond appropriately but ultimately without any feelings whatsoever?
Until the introduction of smartphones, most people would have dismissed a Samantha-like being as implausible. The smartphone is a product of the relentless progress in the computer industry. The emergence of ever more powerful machines replete with ever deeper memory banks is being driven by a hypercompetitive marketplace and by Moore's law, the empirical observation in the semiconductor industry that the number of transistors on a chip doubles every two years or so.
Availability of better hardware has meant that machine-learning techniques invented several decades ago can finally be put to work in software algorithms to allow computers and robots to perform tasks that all of us carry out automatically, day in and day out, without a second thought. These tasks have typically been some of the most difficult ones for computers to execute, but real progress is now being made. Machine-vision algorithms can break down a picture into its constituent parts to identify faces and other objects. Speech-recognition algorithms parse and understand natural human speech. Speech synthesizers take the endless strings of 0's and 1's in which computers communicate and turn them into meaningful speech.
The results are becoming visible outside of computer science laboratories. Machines bested humans in chess years ago, and human leadership in Jeopardy was lost in 2011. Computers translate text from one language into dozens of others and have driven cars over more than 300,000 miles of open roads.
The pace at which machine performance has improved in these early days of the third millennium is staggering—so much so that some pundits predict the imminent advent of true artificial intelligence (AI). It may even be possible to contemplate the arrival in the not too distant future of a digital simulacrum of human-level intelligence. Computers may be getting nearer to receiving a passing grade on the so-called Turing test, conceived by British logician and patriarch of computing Alan Turing in 1950 as a means to discern whether a machine can really think. If a human judge cannot tell whether an answer to a question on any topic came from a computer program or a concealed human, the entity supplying the response must be considered intelligent.
In 1990 Hugh Loebner started an annual competition with a prize of $25,000 going to the first program that fools the (human) judges. So far no team has collected the Loebner Prize. The transcripts of these competitions make for hilarious reading, as human foils try to trip up judges by giving outrageous answers reminiscent of dialogue at a cocktail party for mushroom-ingesting dadaists.
Most academic and industry experts agree that an AI comparable to the intelligence of a typical adult—technology that can learn, infer and generalize in the way we do every day—remains a distant dream. Present-day software can't deal with complex linguistic utterances. It can't figure out that Noam Chomsky's paradigmatic “Colorless green ideas sleep furiously” is meaningless or that James Joyce's “The heaventree of stars hung with humid nightblue fruit” is an eloquent phrase that works its magic by the richness of its imagery. Of course, many people would find these sentences challenging unless they forgo any attempt to apply logic and just delight in the sensuousness of the word pictures, a capability far beyond the reach of any computer.
Setting aside these caveats, machine learning is the hottest technique in the market driving big data analytics. Its practitioners are in high demand, and the technology has been embraced enthusiastically by universities, defense and intelligence agencies, and companies—not just obvious ones like Google, Facebook and Amazon but also Walmart, Target and hedge funds. It remains to be seen whether refining the current crop of machine-learning algorithms will be sufficient for human-level speech or whether fundamental, Nobel Prize–winning breakthroughs will be essential. What is certain, though, is that unlike other standard-fare sci-fi predictions—faster than light warp drive, time travel or radical life-span extension—a Samantha-like verbal intelligence will be born within the lifetimes of many readers of this column.