Volume 2, Issue 1 
1st Quarter, 2007


The Role of AGI in Cybernetic Immortality

Ben Goertzel, Ph.D.

Page 4 of 7

Approaches to AI

AI is an umbrella term – “artificial intelligence”used today to cover a lot of things.  I don't think it's a terribly good term because, after all, an artifice is a tool and AI's may not want to be our tools.  It may not be appropriate. 

In the end if you view the physical universe as a kind of computing infrastructure, the distinction between artificial (i.e. computational) and biological intelligence comes to seem kind of arbitrary.  But I'll accept the word AI because it’s well known; people know what I mean when I use it. 

There are various types of systems that can be grouped under this label of AI.  One type is what I call a narrow AI system, which is I believe, a term I picked up from one of Ray Kurzweil’s books.[1] What that refers to is that systems that are highly intelligent in some narrow domain.  Deep Blue[2], the chess playing program is a good example of that.  Or, the system created by Sebastian Thrun and his team at Stanford University recently which won the DARPA Grand Challenge for artificially intelligent automobile driving.[3]  That's another example. 

These are great programs; I'm very excited about them.  They just do one particular thing.  They don't have any reflective awareness.  They don't have any understanding of context.  But they do one thing very intelligently. 

One of the big lessons in the history of AI research over the last 4 or 5 decades has been the small amount by which progress in narrow AI, actually contributes towards the goal of more general reflective AI.  That wasn't really foreseen.  I think in the 1960's it was thought that making programs like Deep Blue or Mathematica [4] or Google would be big steps toward getting generally intelligent programs that can really think. I don't think it quite works it out that way. 

The narrow AI programs have provided useful tools and insights -- but it's not quite accurate to say they're stepping stones along the way to general AI.  That's a valuable and interesting scientific lesson. 

Another kind of AI system that’s interesting to think about is a totally general intelligence. There's been some fascinating theoretical work done by some European computer scientists, Marcus Hutter [5] and Juergen Schmidhuber [6] and some of their PhD students in Switzerland.  What they have proven, using some really complex mathematics, is essentially: If you have arbitrarily much computing resources, you can get an arbitrarily powerful general intelligence.  That may seem obvious but to prove it rigorously took a lot of advanced mathematics.   

This sort of theory is nice, but I don't think it's terribly useful in terms of making intelligent systems that can do anything.  Because the totally general AI’s they describe require more computing power than exists in this physical universe (by the current estimates).  I think their work is philosophically important, in terms of indicating that the real problem with AI is computational resources. 

The whole challenge of AI is getting intelligent behavior given the limited processing time and memory that actually exist.  If you have arbitrarily much computing resources, their theoretical work shows pretty nicely that you can get arbitrarily much intelligence, according to a pretty general and reasonable mathematical definition of what intelligence is supposed to be. 

Finally, we have the sort of AI which interests me the most, which is what I call artificial general intelligence.



Image 4: Varieties of AI

What I mean is not that general intelligence can do anything, but intelligence that solves more than one problem – intelligence that is able to go into a context and figure out how to achieve its goals in that context autonomously and proactively. 

As an example, rather than just playing one game like a narrow AI program like Deep Blue does, an AGI should be able to learn how to play a game by example and figure out the rules on its own.  It should be able to do what a human can do: Go into a new country, learn the language, learn the customs and figure out how to represent information for itself.  It's kind of a fuzzy definition when you start talking about general intelligence as opposed to narrow AI or totally general intelligence – its level of generality lies somewhere in between. 

Next Page

Footnotes

[1]. Ray Kurzweil - pronounced: [kɚz-waɪl]) (born February 12, 1948) is a pioneer in the fields of optical character recognition (OCR), text-to-speech synthesis, speech recognition technology, and electronic keyboard instruments. He is the author of several books on health, artificial intelligence, transhumanism, technological singularity, and futurism. Wikipedia.org February 8, 2007 1:54 pm EST

[2]. IBM Deep Blue - The computer system dubbed "Deep Blue" was the first machine to win a chess game against a reigning world champion (Garry Kasparov) under regular time controls. This first win occurred on February 10, 1996. Deep Blue - Kasparov, 1996, Game 1 is a famous chess game. However, Kasparov won 3 games and drew 2 of the following games, beating Deep Blue by a score of 4–2. The match concluded on February 17, 1996. Wikipedia.org February 8, 2007 1:56 pm EST

[3]. Sebastian Thrun & DARPA Grand Challenge – Director of Stanford University’s Artificial Intelligence Lab http://robots.stanford.edu/ February 8, 2007 2:00 pm EST

[4]. Mathematica - From simple calculator operations to large-scale programming and interactive-document preparation, Mathematica is the tool of choice at the frontiers of scientific research, in engineering analysis and modeling, in technical education from high school to graduate school, and wherever quantitative methods are used. http://www.wolfram.com/products/mathematica/index.html Fenruary 8, 2007 2:03 pm EST

[5]. Marcus Hutter, Ph.D. – an Associate Professor in the RSISE at the Australian National University in Canberra, Australia, and NICTA adjunct. His current interests are centered on reinforcement learning, algorithmic information theory and statistics, universal induction schemes, adaptive control theory, and related areas. http://www.hutter1.net/ February 8, 2007 2:07 pm EST

[6]. Juergen Schmidhuber - Prof. Jürgen Schmidhuber's main scientific ambition has been to build an optimal scientist, then retire. In 2028 they will force him to retire anyway. By then he shall be able to buy hardware providing more raw computing power than his brain. Will the proper self-improving software lag far behind? If so he'd be surprised. This optimism is driving his research on mathematically sound, general purpose universal learning machines and Artificial Intelligence, in particular, the New AI which is relevant not only for robotics but also for physics and music. http://www.idsia.ch/~juergen/ February 8, 2007 2:19 pm EST

 

 

1 2 3 4 5 6 7 Next Page>