Elusive Intelligence

Upon hearing this kind of scenario about a computer system that carries out our commands, any journalist or lay person immersed in the Internet will jump up and say, “What you really need is an intelligent agent – an electronic servant that will do what you want in your stead.” That’s absolutely true. Such a piece of software would be as natural as having another human being next to us, and it would represent the greatest simplification possible. Give it to me and I’ll throw away every computer system I own and every new system I am designing. The problem is that despite the incessant reference to “intelligent agents,” as if you could go to the corner drugstore and buy them, no one has built them and no one knows how to build them.

A field of research called artificial intelligence was established in 1956 by scientists from Carnegie-Mellon University, MIT, and Stanford University with the twin goal of making machines behave intelligently and understanding how people think. The field is still going strong and has resulted in several innovations now considered in the mainstream of information technology. But the first goal – the injection of humanlike intelligence into machines, known as “the AI problem” – has eluded solution by some of the world’s best scientists and technologists for nearly half a century. And no such solution is discernible on the horizon. No one has been able to imitate by machine the common sense exhibited even by the average toddler.

The “intelligent agents” touted at the turn of the 21st century have been mostly programs that carry out a thin sliver of elementary, humanlike logic via what computer scientists call if-then-else procedures. For example:

If the car phone rings and it the radio is on, then mute the radio.
If a call is initiated and the radio is on, the mute the radio.
If the phone ia hung up and radio is one, then unmute the radio.

If a program like this shows a tiny portion of humanlike behavior, it is dubbed “intelligent,” usually for marketing purposes, reinforcing the illusion that intelligent agents are commonplace. But even the most advanced programs constructed to date in various labs can behave in a marginally humanlike way only in a very narrow context – like Mercury system did for airline reservations. That is very useful for human-centric computing. But it falls far short of the breadth of understanding insinuated by the ambitious term “intelligent agent.” Let’s not fall prey to the syndrome of accepting a wish, stated with a fancy name, as an established capability.

The future prospects for machine intelligence are unknown, as are the fruits of all high-risk, high-payoff research. There was a lot of philosophies, approaches, and beliefs, but no one can responsibly state how far we’ll be able to go toward emulating by machine the intelligent behavior we normally associate with people. That does not diminish the importance of looking for answers. The problem is central and merits more attention that it is getting today, as a result of past disappointments. The kingpin of machine intelligence is machine learning – the ability of a machine to learn from its “experiences,” as it goes along, rather then relying on a human programmer to tell it how to behave intelligently.

With no assurance that machine learning and machine intelligence will happen, we must set aside such wishful thinking and move along with human-centric approaches that will help us interact with our machines naturally – with speech and vision.


Excerpt from “The Unfinished Reveolution” by Michael Dertouzos.