Chris McKinstry has used his knowledge of the Internet, neuropsychology, programming and statistics to single handedly launch the next Apollo program. Well, he doesn't really plan to walk on the moon, but McKinstry feels his giant leap for mankind will be equally significant and will probably require many times the resources of the original Apollo program.
GECC Spring 2001
McKinstry is amassing what will some day be the largest database of common human knowledge on the planet. His project, called MindPixel (www.mindpixel.com), is harnessing the input of thousands of Web surfers, (all of whom will be part owners in MindPixel when it incorporates) to build an ever-expanding record of human knowledge. McKinstry is hoping that with a boundless database, some creative wiring and superfast hardware, his project will give birth to the world's first true artificial intelligence (AI).
Artificial intelligence has long been considered an interesting plot device for science fiction storiesbut not much more. Hopes have been high with people dreaming up applications for AI since the dawn of the computer age, but their expectations have always been low. For the past twenty years, the general consensus among computer scientists has been that modern software and hardware are much too slow to offer us a true HAL 9000, the intelligent computer star of 2001: A Space Odyssey.
The origins of the concept of artificial intelligence can be traced back decades in popular science fiction, but as a true scientific pursuit it is a very young area of study. Most agree its beginnings reside with Alan Turing. In 1950, Turing, who also helped develop the Automatic Computing Engine, one of the first digital computer attempts, wrote a landmark paper now referred to as the "Turing Test." The test called for a human to use a computer terminal to interact in conversations with several different people, as well as with the machine. If the human could not determine which of its conversations were with a person and which were with the machine, the test had been passed and the machine would be considered intelligent-now known as "artificial intelligence," a phrase later coined by John McCarthy, co-founder of the MIT AI lab in 1956.
Since Turing's days, researchers have come to realize that their high hopes for artificial intelligence will likely not happen all at once, but as a result of much smaller steps. In fact, in 1958 prominent researchers H.A. Simon and A. Newell predicted that computers would be able to compose classical music, play chess at a grandmaster level and translate spoken languageall by 1970.
Since then the optimism of AI scientists has cooled off and AI research projects are usually split into two categories, strong AI and weak AI. Strong AI is the development of software that will someday think and reason at, or above, the level of a human being. This is the goal of McKinstry's project. Weak AI, which has already been implemented in everything from automotive systems to videogames, is the development of minute individual intelligent features that enhance a current technology. For example, consider those bad guys in your shoot-em-up video game who always seem to know where you're going to blast next or that infernal paperclip in Microsoft Word who tries to anticipate what kind of document you're about to create.
Because AI is such a young science, and no one has really achieved the holy grail of a computer who thinks on its own, there are many differing schools of thought competing to produce true AI. (See sidebar for some of the types.)
The MindPixel project attacks the problem by breaking intelligence down to its most elemental state-near-binary statements of fact or "mindpixels." At the center of MindPixel's enormous database is a Generic Artificial Conscious (GAC, pronounced "Jack"), which is a software program designed to make sense of the billions of tiny bits of information that will reside in the computer's memory. GAC will form a neural network out of the data and make cross correlations to fill in bits of information that it's not given. Here's an elemental example. If GAC's database contains the following two statements: "Water is liquid," and "the Pacific Ocean is water," then he will be able to deduce that "The Pacific Ocean is liquid."
Years of experience have convinced McKinstry that his current model is the best solution for collecting data and developing AI.
"In 1994 when I first decided to collect this data, I asked people to enter binary items and their responses without any method of validation." McKinstry was able to collect about 450,000 bits of data, but it was loaded with noise. In order to make sure the database would only contain relevant information, and to filter out the absurd, objectionable and false material that was sometimes entered, McKinstry developed a system of validation. Each time a person enters a mindpixel, they are forced to judge and validate or veto twenty other statements.
Through his last ten years of work, McKinstry has come to one startling conclusion: "It is my firm belief that true AI is impossible without the net and that every attempt prior to my 1994 project was comical. True AI is the new space program. It is a giant project that requires giant resources. If you were to try to duplicate the project without the Internet, it would literally cost many times that of the Apollo program."
He notes that GAC is "literally using and needs the entire planet's communications and computing infrastructure" to collect information. A system McKinstry defines as "worth hundreds of billions of dollars."
One other major hurdle that McKinstry faced when he started the project was how to promote his site. "I used everything I knew about branding and marketing to create GAC. I intentionally created a character instead of just a database because people understand characters and how to interact with them. I wanted to give the people entering the data someone to watch evolve over time." Since its launch in July GAC has been written up worldwide from Wired to the BBC Online.
Work in this field is not a new pursuit for McKinstry. He has been interested in developing AI since he was very young. "In 1975, when I was 8 years old, I read an article in the newspaper about two students in my home city of Winnipeg, Canada who wrote chess playing computer programs." McKinstry was smitten with computer intelligence from that point on. "For four years I filled notebooks with ideas to write my own chess playing program and finally managed to do it in 1979. I had to key in my software to a TRS-80 Model 1 in a Radio Shack store as my single mother could never afford one. Since that day in 1975, I have never stopped trying to make computers behave in an intelligent manner, nor will I ever."
While studying honors psychology at the University of Winnipeg, McKinstry was offered a job to develop software for classified military systems. He dropped out of school to take the position and started making more money than his professors. "Within four years I became an executive and returned to school part-time, but still didn't finish. I found it much easier to just buy the textbooks and study on my own. I have some 8,500 books in my library, all read and annotated. I spend six hours per day, every day, reading books."
McKinstry cites statistics, neuropsychology and genetics, in addition to computer science, as the most important parts of his education for developing his career as a pioneer of AI. "Of these, statistics was the most important. The program I was in required graduate level statistics, which I took during my first years instead of my last," he explains.
Uses of AI
When someone does reach the promised land and develops a fully independent and intelligent computer, what will it be good for? Well, chances are you're already dealing with some AI agents in your daily work.
AI allows search engines to do "fuzzy" searches and find material that is related to a topic, but not directly correlated. Once AI searching software becomes more sophisticated, it will be possible to let a preprogrammed AI crawler (a program that moves through Web pages) out onto the Internet and have it return with only relevant information.
Advances in AI will also lead to far more intuitive computer interfaces. This is already happening to some extent in customer support forums and automated telephone systems. Weak AI software allows the computer on the other end of a telephone line to understand your voice commands, so you don't have to continually punch in numbers. Neurostudio, a software package produced by NativeMinds Inc., allows you to create your own Internet support chatter bot. Once you program all of the relevant information about your product into the bot and let it loose on your Web site visitors are able to converse with it as they would a human-instead of spending their time searching through FAQs or using search engines. You can see an example of their technology at www.nativeminds.com.
Why Is AI For You?
Dr. Bart Selman, professor of computer science at Cornell, says the AI job market is potentially unlimited at the larger research labs, developing ingenuity like next generation software. "You need distinct interest in your own brain. Computers are still so close to infancy that any student can help AI grow up. This is more relevant nowadays with Internet start-ups wanting to use information retrieval techniques, thereby offering one way to start working directly out of school." Selman adds, "Even at the most discriminating labs, if you have the skills, there will be a place for you."
Indeed, your bachelor's in computer science alone may be enough to get your foot in the AI door. Just look at a want ad for jobs at the Jet Propulsion Lab:
"...seeking candidates at the B.A./B.S., M.A./M.S. and Ph.D. level to work on fundamental research problems leading to unique software applications in spacecraft autonomy, scientific data analysis, and missions operations automation."
AI has always carried a two-edged sword: make robots life-like and you make progress, make robots life-like and you make trouble. What is an AI creator's responsibility? Can they ensure these machines won't become the nightmare?
"Pressures are on designers to turn out computers that provide better services. If they don't follow through, they'll be out of work quick. More than anything, people's preferences will decide the behaviors in computers that will advance. If they get scary, they will not be used. At least, I hope not," says Eric Horwitz, senior researcher at Microsoft Research and manager of the Adaptive Systems and Interaction Group.
Selman, previously a research scientist at AT&T Bell Laboratories for seven years, believes AI machines should be created with built-in limitations, so humans can have the final say and therefore take the final responsibility. "If we are able to ultimately create fully autonomous entities then clearly we'll need to develop a layer around the entities' software that deals with the issues of safety and privacy necessary to protect humans from eminently more powerful machines."
McKinstry has a more positive outlook on the fruits of his labor. "I hope that people learn to finally see that we are not our bodiesto see we are something independent of physical implementation. I hope we finally learn that we are, more than anything else, information."
Just recently, a robot was created that successfully created another robot within a Brandeis University lab. The breakthrough was that the computer was designed to follow an evolutionary pattern, with each successive robot designed to be better than the first. It was successful. Therefore, while the AI job market continues to expand, one wonders if it will only be comprised of humans. Can you compete with a robot's resume? The bottom line is that AI is not here to replace us, but to understand why we act as we do.
Most of us use approximately 10 to 15% of our brain's capacity. One can only imagine the possibilities if machines could unlock the remaining 90%. If you can imagine, you may find yourself in an unlimited career.
Read more about New Developments in Artificial Intelligence.