Back in 1978 came out the book “What Computers Can’t Do: The Limits of Artificial Intelligence”, by the philosopher Hubert Dreyfus. This book was a strong attack on the Artificial Intelligence discipline of the day; where AI researchers were optimistic that computers will soon attain human-like intelligence because the brain is essentially a computer manipulating discrete data, Dreyfus maintained that this underlying assumption is dead wrong, and that therefore computers cannot and will not attain that goal. Much controversy ensued, of course, but Dreyfus had a point: the AI paradigm of the 1970s was headed towards a dead end. Then, in 1992, Dreyfus issued a revised version: “What Computers Still Can’t Do: A Critique of Artificial Reason”. He updated some details, but the theme remained: computers are limited in acting intelligently like people.
I’m beginning to suspect that in a few years we may see a book created by a computer, titled “Think Again, Dreyfus! We Can Do Anything!”.
I thought of this notion recently while waiting at a traffic light. This light is at the pedestrian crossing at the light rail station near my home, and the only time anyone crosses is after a train stops there. Most of the day the light makes cars stop for no reason. Being an engineer, I immediately started considering solutions: can’t they link the light to the arrival of a train, with timing the system certainly knows? But what if nobody got off? OK: can’t they link it to the presence of pedestrians at the curb, using some sensor to detect them? But then, what if the people are just standing talking, and have no intent to cross? So – can’t the system figure out who plans to cross, the way a human driver would?
And five years ago, that’s where I’d stop, because the behavioral cues of a person about to stop talking and actually cross were too subtle for a computer of that time to discern. It would take a human to do it. But that was then…
Today, I have no doubt that a computer is able to look at a video and decide whether any persons in it intend to step into the street – or if it isn’t today, it will be in a very few years. This is because the developments I’m seeing almost daily in AI, and the rate at which they progress, are growing at breakneck speed to encompass capabilities previously reserved for humans. In fact, my line of thinking goes like this: if a human can figure it out from the video, it means that the information is in there; if it is, then of course a computer can extract it – because a computer can do anything a human can do. Maybe no one bothered to ask it to do this specific task, but if it’s useful, someone will. And no, let’s not start with “Computers will never be able to paint a beautiful painting, or compose a symphony, or fall in love”; the first two tasks are already done, and as for love – I would concede the point if it weren’t that some computers – the ones in our heads – already do experience it. Let’s us wait and see, and meanwhile return to practical tasks that computers are mastering right now.
Every day brings me face to face with such tasks. Until recently these were isolated big deals like winning the world chess title (IBM Deep Blue, 1997), the world Jeopardy title (IBM Watson, 2011) or the world Go title (Google AlphaGo, 2017). But in the past year I notice achievements of a different kind: computers are mastering hard everyday tasks that are much more mundane – and useful. I mean, it’s fine that AlphaGo beat the world champion at Go (a game far harder than chess), but it’s hardly a useful capability. But a computer that can look at a surveillance video and raise an alarm if the people in it are turning violent – that’s a useful functionality with obvious applications. Or a computer that can advise a judge whether releasing a given suspect on bail will endanger the public – that’s already somewhat scary. Or a computer that can read millions of past newspaper stories and learn from them how to predict next year’s epidemics, revolutions and natural disasters (been done). Or the computer that looks at the road in front of my car and alerts me to half a dozen dangers – from potential fender benders to pedestrians jumping into the road. Or…
And the thing is, we don’t even know how these computers do their thing. Those AI researchers in the seventies that Prof. Dreyfus took on were confident they will soon decipher the algorithms used in the brain, and code a computer to be like a brain. But today’s AI relies heavily on machine learning, on big data analysis and artificial neural networks, and these techniques can learn to solve a problem without their programmers ever knowing explicitly how they do it. All they need is a large corpus of training data – of which there is plenty in today’s interconnected information world – and they extract what they need to attack new data. And create new insights; or automate complex processes; or save lives.
And this is today. With computer power doubling every couple of years, in 20 years computers will have 1,000 times the power they now have. I’d be wary what I tell them they can’t do!