- Special Sections
- Public Notices
In recent years the ability of computers has grown dramatically. Many are predicting that machines (computers) will soon be smarter than people. How realistic is this?
If it is a realistic possibility how soon might it happen?
Smart is a poorly defined term in this context, so let’s start with some clarification. For humans the concept of IQ (Intelligence Quotient) is well established and numerous kinds of IQ tests have been devised and used.
An IQ test measures 4 or 5 of the cognitive abilities we humans possess and results in a numerical value that compares us to the average of the rest of humanity. An IQ of 100 means we are exactly as smart as the world average.
Except of course it’s not quite as simple as that since test groups, test structure and testing methodology can possibly skew the results. Rather get lost in a discussion of this topic let’s skip to the related question.
How smart are computers? How fast are they getting smarter? AND how do we measure computer intelligence?
First of all, we still don’t really even know what we mean by computer intelligence! Rather we talk about computer performance. For that there are several well-defined metrics. Well defined, but perhaps not very meaningful in the present context.
The two that are most commonly used are MIPs and FLOPs meaning respectively Millions of Instructions per second and FLoating point OPerations per second.
Much more meaningful metrics are used to compare the relative performance in doing specific tasks. These are collectively called benchmarks and as much as anything now in use allow us to compare human performance to computers.
Then, how fast are they getting smarter? The above metrics show an exponential growth in performance. Performance is doubling every two years or so and will probably continue to do so for at least another 20 years.
The field of artificial intelligence (AI) deals with the possibility and strategy of making machines/computers do tasks requiring some type of cognitive ability.
A relatively new discipline called artificial general intelligence (AGI) focuses on giving machines human like cognitive abilities in many diverse areas.
To expedite and elaborate on the quest for super smart machines, two additional acronyms will be useful. SI (semi-intelligent/specialized intelligence) will be used to denote machines with advanced cognitive ability in a narrow or limited field.
QI (quasi-intelligent) will denote machines that appear to be broadly intelligent (smart), but at least to some extent are faking it!
Both SI and QI are now undergoing rapid and even spectacular development. For the tech savvy, QI is oriented to passing the so-called Turing test. The idea here is that if a human interviewer can’t tell whether it’s a human or machine he’s dealing with, the machine has reached an important milestone.
Another thrust of QI is to design a machine that can take a standard IQ test and get a reasonable score. Both of these goals will be achieved in the very near future yet this will not be at all indicative of a smart machine.
In the field of SI machines, which includes the bulk of present AI activities, special purpose computers are being developed to direct and control all types of machines and equipment. Success in this area has the potential of replacing human workers in ever more disciplines, which could lead to a social crisis at some point in the future if it is not adequately addressed.
True intelligence requires two aspects of cognitive ability that at present are being virtually ignored by AI researchers. One of these, self-awareness, requires that the computer can continuously monitor its own thought process, as well as backtrack and re-examine prior thoughts at random and at will.
This requires associative memory and a totally different architecture than that of today’s computers.
The other is also related to computer architecture and involves developing a process for meaningful learning, in the sense of having and making use of a true understanding of causality, that is, what causes something to happen.
This is an especially tough problem but it is also essential!
All these are soluble problems but probably 20 years or more of research and a few ultra-brilliant engineers will be required. In the meantime the QI folk will fool most of us. By then SI will help machines do much of what only humans can do today.
Art Morse is the president of the Los Alamos Center for Science and Economic Policy.