Terry Brugger

Psyc-392I Fall '95


or The Evolution and State of Artificial Intelligence (AI) [1]

It has long been a dream of humans to create a machine that thinks as they do, if not better. One of the original designers of the modern digital computer, John Von Neuman spent the last few years of his life dedicated to this pursuit. Simon and Newell who have had one of the greatest successes in the area with their General Problem Solver (GPS) proposed that computers would "be capable of composing classical music, discovering important new mathematical theorems, playing chess at grandmaster level, and understanding and translating spoken language." This paper will briefly examine the current state of AI, its related and component fields, and some solutions to current limiting factors, which naturally leads to a proposal for a new AI system.

Intelligence is a fairly new idea. Life has taken some 3.5 billion years to develop on Earth, and what we know of as intelligent life has only come onto the scene in the past few ten thousand years. If you were to look at how quickly humans have evolved in that time, the results are astounding. Hence the theory of many computational scientists is that breeding grounds for instruction sets, which define the characteristics of a computer as DNA define the characteristics of a human, be developed where these instruction sets can replicate and, in the process, mutate. Over time, strong and viable instructions will result. Those instructions will begin combining into instruction sets, much as one celled creatures developed into multi-celled animals. From here, the instructions' complexity will jump through the roof until, a relatively short time later, they develop intelligence on their own. Many projects have shown the viability of this; the most popular of these is Tierra, which has demonstrated instructions which can develop immunity against parasites. There is a proposal in the works to create a world wide habitat over the Internet in which these creatures can live and develop.

Of course, most people don't believe in reinventing the wheel. Perhaps we should just use the intelligent computer that nature has already created. Hence neural networks have gained popularity in the past decade as tools that can be popped right out of the box and start learning to make decisions right away. For instance a neural net can be taught what products are good or bad, and then be used for quality control on a production line. While the neural net is probably the most useful thing to come out of AI research, there is still a major limiting factor to its usefulness to AI -- the sheer size and complexity of an intelligent brain. While there are artificial neural networks in use that use up to hundreds of thousands of neurons, they are still magnitudes short of replicating the billions of neurons in an intelligent brain. Even if a network this large could be implemented on today's computers, there is still too much that we don't know or understand about the synaptic connections in the brain to be able to properly simulate it.

From here we jump over millions of years of evolution from the neuron to intelligence. Intelligent agents are computer programs that are designed to assist humans by use of the agent's artificial intelligence. There are two major classifications of agents: personal and business. A personal agent is one that a person would use that "learns" how the user works and provides assistance when it can. The biggest hope for personal agents at this time is that they will be able to learn what you do and do not like and be able to filter information that you receive (e.g., recommend television shows instead of having you channel surf or have to look up the listing)[6]. The business agent is one that allows people and companies to work more efficiently. An example of this is the CD finder[7] operated by Anderson Consulting. The CD finder takes the title and artist of a CD that you specify and contacts 9 music stores on the WWW to find the lowest price for you. The success of these tools has come under debate as of late. Some argue that personal agents are just oversimplifications of the user interface, at the expense of control by the user. The publicly available agents such as CD finder do not appear to demonstrate any intelligence. The agent appears to just be a script that logs onto each music store's site and look for the album. In a search for The Cure's Wish, it was only able to find two stores with copies. It reported that it had trouble finding it at three others, was being refused service at three others, and that WWW service had been discontinued by another. If it was truly intelligent, it should have had a better idea if the album did indeed exist at the three stores in question. It should have had a better idea if the album did indeed exist at the three stores in question. It should have found an alternative way to get into the three stores that refused it service, or more precisely, it should seem intelligent enough that the store doesn't refuse it service in the first place on the basis of being an agent. Finally, it should not be limited by the nine stores programmed into it; rather it should constantly be on the lookout for new music stores and dynamically eliminate those that discontinue service.

Other AI systems have shown a remarkable level of coherent intelligibility. These systems are primarily designed to pass the Turing test. Turing proposed that an AI system that could communicate with a human, without that human knowing that they are talking to an AI, had succeeded in attaining intelligence. Some systems such as Eliza were designed in the late 60's and early 70's that are very accurate in communicating with humans in a limited topic area. Eliza, for instance, shows remarkable intelligence in a conversation where she acts as a clinical psychological therapist. Modern AI systems have been programmed with more areas of knowledge, e.g., Colin[8] is programmed with a rudimentary knowledge of computing; when Colin is asked, "What is artificial intelligence?" he enthusiastically replies, "I know a *LOT* about AI!" In a search for more details, he was asked, "What is the greatest limiting factor in AI research?" to which he bluntly replies, "all the researchers just like! to publish bullshit papers." At least the greatest limiting factor is not their vocabulary.

Expert systems lie somewhere in the middle of all this research. Expert systems are computers that have been programmed with the domain knowledge of an expert in the field. The system contains heuristics for solving certain types of problems, such as medical diagnoses. It is in this domain specificity that the systems are most limited. They are further limited because they are programmed with just the knowledge acquired from one (rarely a few) expert in the field, and lack much room to grow or develop. Nonetheless, expert systems are claimed by many to be the most useful things that AI research has developed.

If these other "intelligences" fail to demonstrate some level of cognitive ability, then surely there must be some AI system out there that can. There is only one major system that has shown promise of true cognitive ability. This is the CYC[9] system. The CYC system was first started as a ten year project in 1984. It was extended another decade in 1994. While the initial problems of setting up an AI were relatively quickly surmounted, CYC's development was still limited by its database of simple knowledge. This database consists of facts like, "Most creatures with two arms have two legs." The hope is with a huge knowledge database like this, that CYC will become a general expert system. CYC has shown promise as a viable, commercial product, and implementations of it are expected to be put into use in the corporate world within the next year[10].

If CYC is successful in implementing a machine with a great knowledge base that can think on its own, what's to stop it from being essentially human? Many argue the unique human trait: emotions. However it would seem that if intelligence, which at the basic level is just neurons firing, can be abstracted to a form a computer can emulate, then emotions, which at the basic level are the same thing, can be as well. In fact the basic research in this area not only predates digital computers, but was done by a literary author. James Joyce's Ulysses is the story of a common man whose emotions have been abstracted into their component parts for literary purposes. However, this abstraction lends itself to a framework that can be relatively easily implemented on computers so that they may not only think, but feel as well. As human nature has taught us, emotions can be quite strong. Dan Lawrence, author of MicroEmacs[2] believes that the most useful emotion to computers will be motivation. If computers have a true motivation to do something, they will use the resources available to them to get it done, much as humans do. As discussion on factors such as emotions explore further in depth, they quickly become philosophical: What if a computer feels motivated to control the world? Questions such as these are beyond the scope of this paper and are better left to AI's such as Deep Thought II to ponder [3].

So it would seem that a useful Artificial Intelligence is a viable prospect, so why are computers still so stupid? Well, it would seem that the greatest limiting factor in AI research is indeed "bullshit". Each of the above systems is based on its own ideas and the authors seem to do little but criticize other approaches to the problem rather than admit that another approach lends something useful to the AI as a whole. Alife seems like a promising prospect, but at the expense of time. To speed up its development requires pure computational power. This would best be approached by support of the proposal for an artificial life reserve. Neural networks similarly are a viable prospect, but again at the expense of computational power. Intelligent agents have far to go to demonstrate any intelligence. Parsing AIs like Colin and Eliza are great to talk to, but like intelligent agents, they lack any real intelligence. Even worse, they have very little data to go by. Finally, the great 20 year CYC project, which many believe holds the greatest prospect for a tangible AI, must be spoon-fed information on a daily basis. Hence the following AI is proposed: The basic search algorithms located in intelligent agents and the front end of Colin or Eliza to communicate with humans, or more spectacularly, each other. Finally, it could be given emotions, which would aid it, just as human emotions such as motivation increase our performance.

The implementation of such a machine is almost incomprehensible. Even with the state of electronics miniaturization today, it would take up several rooms just as the first computers did. This is infeasible. Rather, the AI should take advantage of the computational resources available on the Internet. If the core engine of the AI is written in a dynamic massively parallel language such as Piranha, [11] the core engine could be started on a small system, such as a lab of Sun SparcStations[12]. The AI would then begin spreading to other computers, just as a parasite spreads[13]. At first it could easily exploit major holes in security existent on countless machines located on the Net. Since the AI would run as a background process on these machines[14], it would only execute during times that the machine is not being otherwise used. As a result, no user should notice a change of performance in their computer and unless someone searched for it, the AI would not be detected[15]. As its computational power increases by the inclusion of more and more nodes, it can attempt to crack more complex systems. Eventually, it should be able to implement itself on increasingly powerful machines, particularly the massively parallel research supercomputers like the Touchstone Delta, currently the worlds fastest computer[16]. Obviously, like a child, it will make mistakes such as trying to crack into the US Defense Department's computers -- as attempts are made to kill it, its motivation to stay alive will teach it not to do stupid things like that. The other limiting factor in its development would be its knowledge database. Would a system as smart as that not require an immense amount of storage to keep all its knowledge in? Well yes, but this storage already exists -- the Internet on which it exists. The actual knowledge of its existence and what it's doing will just be passed as packets between nodes, like neurons firing. Its database of facts will be likewise spread throughout the net -- its basic knowledge will include how to find information, much as this author did in researching this paper. As it grows, it will develop more efficient ways to look for data. It will quickly become so fast that it can find requested data for human users more efficiently than other currently available searches.[17]

Finally, it should be noted that what has been proposed here is not only illegal in this country[4], but very dangerous as well. As stated earlier, there is the philosophical question of what happens if the AI becomes bent on domination? Simply, there should be a code which when distributed over the network, would order the Piranha host upon which the AI is running to close down, effectively killing the mind's body.

When one looks at the current state of AI systems, it's easy to see how varied and diversified the research is. While exploring every aspect of the field is good, the lack of application of developments in one part of the field to another is somewhat despairing. Hence the system proposed here, Ulysses, would take advantage of the most promising aspects of each of the fields to produce a living, computing, thinking, yet artificial intelligence.

[1] This paper was generated almost entirely electronically. All research was conducted on the WWW and the paper was composed and edited with Word. The point of this is to prove that an AI can exist entirely in the electronic domain, gathering all the information it needs and communicating with the world over the Internet.

[2] Emacs is a text editing tool whose roots lie in LISP, the same language used almost exclusively for AI research up through the end of the last decade. MicroEmacs is a specific variant of Emacs designed for portability across multiple computer platforms.

[3] Deep Thought II holds the current record as being the best AI system for Chess - it ranks somewhere in the range of 100th - 150th place among the world champions. Deep Thought II was also the name of a fictional AI in Douglas Adams' Hitchhiker's Trilogy which contemplated the question, "What is the meaning of Life the Universe and Everything?" [5]

[4] Under federal law, cracking computer systems is a crime. If this crime takes place on a network such as the Internet which crosses state lines, it is enforceable by the Federal Government (which is why many crackers are spending time in federal prison). However the Internet extends to almost every country on Earth, most of which don't have such laws. There would be nothing the government could do to prevent such an AI system from starting in such a country.

[5] Since this paper was written, IBM's research division has developed a new AI system, Deep Blue which comes closer to "Inteligence" primarily through raw computing power.[18]

[6] Since this paper was written, TiVo has developed software for a digital video recorder (like a VCR that records programs to a hard drive) that, among other things, predicts what programs you'll enjoy (based on your ratings of programs you've seen) and records them for you on any available space it may have.

[7] Since this paper was written, CD Finder apparently became Bargain Finder before evolving into Pocket BargainFinder. The agent itself is no longer accessable to the public. It does appear to still exist at it's old location.

[8] I am not able to find a current link to Colon or any information as to what happened to him. Here is the old link to Colon. As best I can tell, Colin was just a cgi (Common Gateway Interface) script to interact with the Maas-Neotek (A reference to William Gibson's Mona Lisa Overdrive) chatterbot developed at CMU, of which Colin was the default name. I have not been able to find any references to Colin, Maas-Neotek or the server that they used to reside on at CMU however.

[9] I now reference the CYC homepage, which didn't exist as of the original writing of this paper. Here is the updated link to the FAQ that was formerly referenced.

[10] Since this paper was written, CYC now has products based on their research.

[11] When this paper was first written in 1995, it looked as though Piranha may hold some promise as a parallel programming language. Unfortunately, as most projects academic, development on it has floundered (no pun intended). In today's environment, I believe the best solution would probably involve a combination of languages such as Perl, Python and a variety of native scripting languages, such as sh, VBScript and AppleScript.

[12] When this paper was first written, a lab of Sun SparcStations was considered a fairly good facility. One might think that this should be updated to read a lab of Sun UltraSparcs, or perhaphs a Linux Beowulf cluster, however I will argue that clusters of Sun SparcStations, while not as advanced as today's offerings, have not slowed down in the past five years. An instalation of SparcStations will probably not be as closely monitored as a lab of state of the art machines. This may facilitate any illicit or unauthorized activities described herein[4].

[13] I neglected to attribute a reference to this when it was first written. What I am refering to is, of course, commonly known as a virus.

[14] More properly, it would run as a background process at one of the lowest priority levels so that it would be run instead of the idle thread/process.

[15] In fact, on most systems it could hide its existance almost altogether through the use of a Rootkit.

[16] Obviously, the world's fastest computer has changed quite a bit since this paper was written. At the moment, it's a toss up between any of the three main ASCI machines: Red, Blue Pacific or Blue Mountain, depending on the application. In the near future, White will easily surpass all three. The reader is encouraged to look at the current list of top supercomputers, based on the performance of Linpac. It is worth noting in the given context that these computers are well protected and attempts by a virus to crack into them would be noticed and not well received.

[17] Since this paper was written, there have been numerous incidents of viruses/worms acting in a primitive matter as described here. The first major incident was the Melissa virus/worm of 1999. While not very disruptive (except for flooding email systems), it was an effective proof of concept. A slightly more insiduous and destructive strain called the I Love You virus/worm spread around the Internet slightly more than a year after the Melissa incident. The "I Love You" worm was followed shortly by numerous mutations, some of which are much more destructive. These viri/worms just begin to scratch the surface of what's possible. A former CIA director has proposed that such worms will be used for covert data collection. Michal Zalewski has posted a paper wherein he describes a much more advanced worm which sounds uncannily like Ulysses in concept.

[18] Deep Blue defeated World Champion and International Grandmaster Gary Kasparov 3.5 to 2.5 in a six game match 5/3/1997 - 5/11/1997 through a combination of pure computational horsepower, specialized hardware and a highly developed and refined expert system. Check out the IBM website with complete details of the match.

Finished: 12/4/95

Modified: 12/8/95 - Fixed links

Modified: 12/9/95 - Fixed links

Modified: 1/17/97 - Fixed missing quotes and added footnote 5

Modified: 5/21/2000

Last Modified: 5/22/2000

This paper is copyrighted by Terry Brugger.