When Joe Weizenbaum found his secretary using a computer program he had created, he was so upset he devoted the rest of his life to warning people not to use its technology. The program was “Eliza”, which gives a passable imitation of a nondirectional psychiatrist; you type sentences such as: “I wonder what I should write,” and it replies :“What answer would please you the most?” (You can try a version at psych.fullerton.edu/mbirnbaum/psych101/Eliza.htm).
Weizenbaum’s distress came because he had written Eliza as an experiment, to see whether he could simulate “artificial intelligence” in a question-and-answer system by parsing sentences and throwing relevant bits back at the questioner. But his secretary saw it as real, and asked him not to intrude on “sessions”; Weizenbaum saw this as an omen that we would be too easily fooled into trusting machines.
The story dates from the 1960s (Weizenbaum died in March 2008), but is relevant today. Machine intelligence and machine learning – the new synonyms for “artificial intelligence” – are on the rise and are going to be pervasive. To some extent, anyone using a smartphone is already using some sort of machine intelligence with Google Now’s suggestions or Apple Maps’s determination of what counts as your “home” and “work”, or Windows Phone’s Cortana. We don’t call these “artificial intelligence”, because the moniker earned ridicule down the years. But it doesn’t matter what you call it; the ability to get computers to infer information that they aren’t directly supplied with, and to act on that, is already here. Take Google Photos, released in May, which can figure out that you were taking photos of a zombie parade, or find pictures of canoes, or - well, pretty much anything. It’s not limited to human faces (though it can do those, and distinguish between children and babies). If you don’t want to know about deep-learning neural networks, then it satisfies Arthur C Clarke’s dictum: “Any sufficiently advanced technology is indistinguishable from magic.”
Part of the magic behind Google Photos comes from a British company, DeepMind Technologies, which built what it calls “artificial generalised intelligence”. They trained it on video games such as Space Invaders and Breakout; it got amazingly good at them after a few hundred games, without any special training. DeepMind isn’t the only machine learning company Google has acquired; last August it bought Jetpac, which used neural networks to figure out what was in your holiday photos.
The promise of such machine learning is hugely exciting, and not limited to Google. Jetpac’s creator Pete Warden wrote a fun app called Deep Belief which you could download for iOS that used neural networks to try to identify what the camera was seeing. . There’s an open-source machine learning framework called Seldon. Another AI system called Snips now in beta (http://snips.net/beta), aims to put a personalised AI – a “smart assistant” – on your phone.
It’s that latter idea that’s the real promise. Eliza used to require a mainframe; modern machines can run it as an applet without breaking sweat. Deep Belief shows that neural networks can run on modern smartphones. Give it a few years, and your smartphone’s AI will have power enough to understand the context of what you’re doing, figure out how it should help (call or text people to say you’re late? Understand whose emails you do and don’t want to receive at particular times? Find the right people for you to link up with on social media?) and act on it. You’ll be able to choose whether you want that done by an “assistant” on your phone, or in the cloud. The cloud-based one will be smarter, because it will have access to much more data; the phone-based one will be much more personal, and function even without a network connection.
Weizenbaum feared that AI couldn’t be trusted with decisions because it wouldn’t have empathy or compassion. But that’s fine. It’s the everyday grind that AI can smooth. I, for one, welcome our new robot underlings.