This was from the first newsletter. Note that the current AI represents a vastly superior way of pattern matching, down to the level of pixels in art work (which is why it requires so much computing power). It has since conquered the game of Go and
I always loved the way Hyperion referred to a historical event (in it’s fictional Universe) called “The Big Mistake” which caused humans to have to leave Earth. Spoiler alert — It was AI.
Welcome to the Premier Issue
Welcome to Dr. Dobb’s AI Expert Newsletter. Before you begin to talk about Artificial Intelligence (AI), you have to deal with the name itself. It leads to impassioned philosophical debates, like those raging right now on the AI Forum, and inspires some great science fiction. (The Hyperion Trilogy by Dan Simmons is my favorite, based on Tom Ray’s fascinating Tierra project, but that’s for a later newsletter.)
Debates on machine vs. human intelligence are fun, but this newsletter needs a working definition of AI that more realistically encompasses the wide variety of fascinating software and hardware usually classified as AI. So we’ll define it in terms of computer nature rather than human nature:
Artificial Intelligence (AI) is the art and science of making computers do interesting things that are not in their nature.
What is a computer’s nature? Well it has memory that’s good for storing data and a CPU that processes sequences of instructions. If you have a problem that requires automating data storage, procedures or both i.e. data processing, then computers are perfect. Other types of problems are trickier.
Consider Mycin, one of the earliest AI programs. The developers were posed with the problem of trying to automate medical diagnostic knowledge that was best expressed as rules.
The rules were neither data or procedures. Trying to force them as either was untenable for any significant body of knowledge. So what could they do?
The answer lies in the words of some long forgotten hardware engineer:
“Software is a kludge for getting around design flaws in the hardware.”
What the Mycin developers really needed was a new kind of computer: one that processed rules. So they created a virtual machine, they called it an inference engine, whose nature was ideal for processing knowledge expressed as diagnostic rules.
People in the AI field have been doing the same thing ever since, creating virtual machines that are programmed in non-conventional ways, and using those machines to perform previously untenable programming tasks.
Which leads to the problem with an AI newsletter. There isn’t a common language, like Java or C++ that is used for AI applications, nor are there standards like SQL for database access, or HTML for Web browsers.
Nor are these virtual machines similar. The classic ones deal with a wide variety of rule languages and reasoning algorithms, while others deal with semantic networks, neural networks, ontological knowledgebases, genetic algorithms, artificial life, robot brains, cellular automata and, of course, games.
And then, given a specific area of AI one can write about how to build the tool, or write about applications of the tool. The former is solid programmer stuff; the latter requires some knowledge of the specific tool, which is often a commercial product.
So I’m asking for your input. There is a wealth of fascinating subject matter that this newsletter can cover, and I’d like to know what your interests and experiences are. Send me e-mail. Tell me what you’re currently doing with AI and what you’d like to see covered in the AI Expert Newsletter. I want the newsletter to meet your interests.
News
It’s a Tie!
Brains in Bahrain was the latest man-machine chess showdown of champions, pitting Kramik (a human) against Deep Fritz (a software program). Kramik won the early matches in what was called a triumph of anti-computer strategy. Deep Fritz won the later matches in what was called a triumph of anti-human strategy.
Chess playing has often been considered an ultimate test of machine intelligence because it is too complex to be mastered using straight-forward search algorithms. That is, the search space of possible moves is too great to be explored by a machine.
Early researchers worked on encoding the essence of grandmaster reasoning to improve performance. They learned, for example, that grandmasters had greater “pattern vocabularies” than weaker players. They didn’t reason faster than weaker players, but more efficiently because they were reasoning using larger chunks of knowledge.
With the speed of today’s machines, the value of the brute-force search in chess had become more important. Its ironic that in this classic test of machine intelligence, computers are playing on a par with the best humans more by exploiting the natural strengths of computers, and less by the application of human-like intelligence. Intelligent or not, today’s top chess programs are masterpieces of software engineering.
It is interesting to note that Deep Fritz is not Deep Blue, which was dismantled after defeating Kasparov in the last man-machine chess championship. Deep Blue was an amazing mix of specialized parallel processing hardware and software capable of searching 200 million moves per second. Deep Fritz is a souped up commercial product from ChessBase running on stock hardware capable of only 3 million moves per second. They claim it’s smarter though.
As for humans, with their feeble CPUs, they still provide stiff competition for millions-of-moves-per-second machines. It all appears to revolve around superior recognition and processing of patterns. And when you take a heavily pattern-oriented game like go, well computers aren’t even close, yet.
1738 Artificial Duck
Dr. Jessica Risken of Standord University has done a study of a mechanical duck built in 1738 that created the same interest and excitement as current experiments in artificial life. The parallels between their desire to use mechanized models to understand real life, and ours, are fascinating.
Among other things, the duck was realistic in that it ate and pooped. But if you looked inside, it sort of cheated. Food went in, and time passed, and poop came out, but the input food was still there, and the poop had been pre-loaded. So the duck didn’t really digest, it just looked like it did.
That is not dissimilar from the chess programs that look like they understand the patterns on the board, but instead are processing millions of individual moves.
But also like the chess programs, the duck was an amazing bit of engineering and probably a lot of fun to build.