Beneath the magical aura of artificial intelligence lies a mundane, dirt-cheap predictor
Typing “technology indis…” into Google instantly directs one to a web page discussing Arthur C Clarke’s third law: “Any sufficiently advanced technology is indistinguishable from magic.” The science fiction writer’s aphorism was published in 1962, when Google’s autocompleting search engine would indeed have seemed like sorcery.
There are many other examples: electricity, aircraft and the telephone would all have seemed miraculous and inexplicable to earlier generations. Each of them exemplified what a technological breakthrough is supposed to look like.
We need to be careful, however, not to overlook much simpler technological advances. The light bulb is a safer and more controllable source of artificial light than the candle or the oil lamp, but what really makes it transformative is its price — the cost of illumination has fallen 400-fold in the past two centuries. Supercomputers and space travel get all the press. Merely being cheap doesn’t. But being cheap can change the world.
Consider barbed wire (cheap fencing), the shipping container (cheap logistics) or the digital spreadsheet (cheap arithmetic). Ikea gave us cheap furniture, and the same principles of simple modular assembly are giving us cheaper solar panels too.
My favourite example is paper: the Gutenberg press radically reduced the cost of producing writing, but it was of little use without an accompanying fall in the cost of a writing surface. Compared with papyrus, parchment or silk, one of paper’s most important properties was that it cost very little.
What therefore are modern technological advances that we may be overlooking or misunderstanding because they are cheap rather than magical? The obvious answer: sensors. We are surrounded by inexpensive sensors — in phones, increas- ingly in cars — continually taking in information about the world.
A new book suggests a different, albeit related, answer. Prediction Machines by Ajay Agrawal, Joshua Gans and Avi Goldfarb argues that people are starting to enjoy the benefits of a new, low-cost service: predictions. Much of what is called “artificial intelligence”, say the authors, is best understood as a dirt-cheap prediction.
Predictions are everywhere. Google predicts that when I type “technology indis…” I am looking for information about Clarke’s third law; Amazon makes a prediction about what I might buy next, given what I have bought already, or searched for, or placed on a wish list. A prediction may literally be a forecast about the future or more generally it may be an attempt to fill in some blanks on the basis of limited information.
Not all such predictions are very good, but not all of them need to be. The tiny keyboards on smartphones turn out to be quite serviceable when combined with modestly accurate predictions — from suggesting an entire one-phrase e-mail reply (“I agree with you”) to subtly expanding the “H” and shrinking the surrounding keys on a touchscreen if the phone thinks that “H” is the more likely target for a fat-thumbed typist. Errors in predictive text tend to be trivial and easy to correct, so a high error rate does not matter much. Clumsy text predictors can be released into the world so that they may learn. A high error rate in a self-driving car is not so easy to forgive.
As Agrawal and colleagues point out, sufficiently accurate predictions allow radically different business models. If a supermarket becomes good enough at predicting what I want to buy — perhaps conspiring with my fridge — then it can start shipping things to me without my asking, taking the bet that I will be pleased to see most of them when they arrive.
Since good predictions reduce uncertainty, there may in future also be less demand for things that help people deal with uncertainty. If that conspiratorial fridge can arrange just-in-time delivery of meal ingredients by predicting requirements, it can be much smaller as a result.
The airport lounge, a place designed to help busy people deal with the fact that in an uncertain world it is sensible to set off early for the airport, is another example. Route planners, flight trackers and other cheap prediction algorithms may allow many more people to trim their margin for error, arriving at the last moment and skipping the lounge.
Then there is health insurance; if a computer can predict with high accuracy the chances of getting cancer, then it is not clear that there is enough uncertainty left to insure.
All this seems a useful way to look at the fast-changing world of machine learning — more useful than pondering Clarke’s most famous creation, the murderous computer HAL 9,000.
Some automated predictions are already marvellously good, but many are changing the world not because they are omniscient but because they are good enough — and cheap. /©
WE ARE SURROUNDED BY INEXPENSIVE SENSORS THAT ARE CONTINUALLY TAKING IN INFORMATION