This weekend’s New York Times has a fine article on AI, The Great A.I. Awakening. I commented briefly on the article on the Times’s site but I want to say a lot more.
The article vividly documents Google Translate’s recent revolution in how it works. Until now, auto-translation engines have modeled languages explicitly via rules, dictionaries, and the like. The new Translate, and its Chinese competitor on Baidu, instead enable a multi-layer neural net – a simulated brain, basically – to learn language translation by being fed thousands or millions of existing examples of translations. Researchers fed Google Translate the complete English and French versions of the Canadian Parliament’s proceedings, for instance, presumably along with many translated classic books, newspapers, and so forth. The new engines learn like human toddlers do, by unconsciously copying behaviors they observe, over and over again, until they evolve to proficiency. And like humans, the new engines will continue learning their entire “lives” by observing and copying new examples with new words and new phrases in all languages. But note: the new engines will not be able to go out in the world themselves to find worthy new examples, not for a very long time yet. They’ll need human care and feeding for the foreseeable future.
From near the end of the article:
A neural network built to translate could work through millions of pages of documents of legal discovery in the tiniest fraction of the time it would take the most expensively credentialed lawyer. The kinds of jobs taken by automatons will no longer be just repetitive tasks that were once — unfairly, it ought to be emphasized — associated with the supposed lower intelligence of the uneducated classes. We’re not only talking about three and a half million truck drivers who may soon lack careers. We’re talking about inventory managers, economists, financial advisers, real estate agents.
All true. But all decades out in the future, maybe several or many decades. Why? I see it like this. A toddler computer learns from a team of humans whose only job today is to feed and raise this child quickly to do one thing well. The toddler computer has no eyes, ears, nose, mouth, hands, legs or feet. The toddler computer processes fed-in data 24×7 and learns its one thing quickly in its tiny simulated brain, much faster than a human child would. But a human child processes many orders of magnitude of far, far richer data per time period than the computer child can: visual, auditory, tactile, olfactory, the physics of standing and walking and making sounds, the complexities of language, and so forth, all integrated and organized within its real brain. See for example this discussion.
The human child’s brain is the culmination of millions of generations of evolving, increasingly more powerful prototypes equipped with extraordinarily capable sensors of several kinds. The human brain perceives and processes the world around it continuously at a huge data rate. The human body moves freely in space. With respect to attaining human-like intelligence and self awareness, then, the computer toddler has a truly enormous gap yet to cross. The crossing cannot possibly be quick, meaning in just a few years or a decade. It’s been only a few years since the total computing power on the planet exceeded just one human brain’s computing power.
Not to diminish or underplay the Google Translate achievements in any way. They are stunning. But I view them as like Watt’s invention of rotary steam motion in the late 1700s: an enormous enabler of a revolution, but still just the very beginning. And I’m not one bit worried by the article’s conclusion that “once machines can learn from human speech, even the comfortable job of the programmer is threatened.” No, not for a very long time yet to come.
As the article says, “The goal posts for ‘artificial intelligence’ are thus constantly receding.” Each step seems major, and Google Translate’s awakening is indeed major, but it’s still tiny in the big picture of true intelligence and self awareness.