TCS Daily


Where I'd Bet Against Kurzweil

By Arnold Kling - August 17, 2005 12:00 AM

"I proposed a bet to Kurzweil. Under the bet I would give him a very small amount of money today and in return at some future agreed-upon date he would give me a 10-meter-diameter solid diamond sphere."
-- James D. Miller

In my opinion, James Miller is making a bad bet. If you want to bet against Ray Kurzweil, you should look for patterns of prediction errors. As this essay will show, Kurzweil has been systematically overly optimistic concerning some forms of artificial intelligence. But in forecasting developments based on brute computation and nano-scale engineering -- which are what is needed to solve the problem of fabricating a large, perfect diamond -- Kurzweil has been, if anything, a bit conservative.

 

Back to Kurzweil's Future

 

In anticipation of Kurzweil's forthcoming The Singularity Is Near, I returned to his 1999 book, The Age of Spiritual Machines. In that book, he made predictions for several subsequent decades. Given that about 6 years have elapsed since the book's publication, I figure that we should be far enough along to recognize which of his predictions for the year 2009 (chapter 9) are likely to come true, and which are likely to go awry.

 

First, we need to remember nonlinear thinking. Suppose that a technology is doubling its capability every two years. Then, if you make a prediction for the year 2009, we should be half way there by 2007 and one-fourth of the way there in 2005. We should have been only 1/32 of the way there in 1999.

 

Any development that was already more than 1/32 in place when Kurzweil wrote ought to count for less in his favor, because a prediction of an improvement by a factor of less than 32 is not so dramatic. For example, Kurzweil's prediction that in 2009 "Military conflicts between nations are rare, and most conflicts are between nations and smaller bands of terrorists" might seem to be a remarkably accurate prediction. However, in my opinion, we were at least one quarter of the way there in 1999. From a nonlinear perspective, this prediction was too easy. I would say the same about the prediction that "Learning is becoming a significant portion of most jobs."

 

Leaving out the too-easy predictions, the typical prediction that Kurzweil made in 1999 ought to be about 1/4 of the way to coming true by now. Using that standard, here are what I see as his most conservative and most optimistic predictions in various categories. If we seem more than 1/4 of the way there now, then he was conservative. If we seem less than 1/4 of the way there now, then he was optimistic.

 

Category

Conservative

Optimistic

Computers

"Computers routinely include wireless technology to plug into the ever-present worldwide network"

"The majority of text is created using continuous speech recognition"

Education

"Students of all ages typically have a computer of their own"

"students can learn basic skills such as reading and math just as readily with interactive learning software as with human teachers"

Disabilities

"Computer-controlled orthotic devices have been introduced. These 'walking machines' enable [some] paraplegic persons to walk and climb stairs."

"A blind person can interact with her personal reading-navigation systems through two-way voice communication"

Communication

"Users can instantly download books...radio, movies, and other forms of software to their highly portable personal communication devices."

"Translating Telephone technology (where you speak in English and your Japanese friend hears you in Japanese, and vice versa)"

Business

"Most purchases of books, musical 'albums,' videos, games...do not involve any physical object, so new business models for distributing these forms of information have emerged."

"Intelligent roads...Once your car's computer guidance system locks onto the control sensors on one of these highways, you can sit back and relax"

Music

"The creation of music has become available to persons who are not musicians."

"Interactive brain-generated music, which creates a resonance between the user's brain waves and the music being listened to, is another popular genre."

Medicine

"Doctors routinely train in virtuality environments, which include a haptic [tactile] interface."

"Doctors routinely consult knowledge-based systems...which provide automated guidance"

 

Tools vs. Agents

 

In comparing the predictions that turned out to be conservative with the predictions that turned out to be optimistic, I detect a clear pattern. Generally speaking, the more open-ended the problem and the more adaptive that the machine needs to be to provide a solution, the less far along we are in arriving at a technological solution. One way to put this is that we can construct tools, but we cannot construct agents.

 

A tool is a device that performs specific, easily predicted behaviors in response to a given set of possible stimuli. A tool takes information from a human, rather than directly from the environment. A human must adjust to provide "input" to the tool, rather than having the tool respond to natural human behavior. A tool follows a well-specified set of steps in formulating its response to stimuli. The tool's behavioral rules can be reverse-engineered relatively easily.

 

An agent is a device that can adapt to new stimuli and produce unexpected responses. An agent can take autonomous readings of the environment. It can interpret natural human behavior. An agent adapts and changes its behavior. Its behavioral patterns are difficult to reverse engineer.

 

I doubt that one can draw a clear dividing line between tools and agents. There probably are some devices that might be difficult to classify. However, at the extremes, the distinction seems reasonably easy to draw. Downloading music wirelessly to a personal communication device requires a tool. Creating a resonance between the user's brain waves and the music being listened to requires an agent. A virtual-reality cadaver for a doctor is a tool. A real-time medical consultation assistant would be an agent.

 

Today's cars are complex tools. That is, a car is a tool that is made up of many individual tools. However, the car is not an agent. The driver still does most of the work of processing environmental stimuli and adapting to the situation. Kurzweil's intelligent highway, in which the car acts as an agent, seems to me to be a lot farther in the future than 2009.

 

A keyboard is a tool. A speech-recognition system is an agent. Speech recognition has been a "promising" and "improving" technology for decades, but it perpetually falls short. I used to call it a Red Sox technology, except that as of last October the Red Sox no longer are a metaphor for championship-level futility.

 

The Learning Factor

 

As I wrote in On Intelligence, People, and Computers, human beings far exceed machines in terms of learning capacity, because we develop generic pattern-recognition skills in response to our sensory environment:

 

"I think that the key to getting any machine to learn is to give it a variety of both stimuli to absorb and tasks to perform. Moreover, it is important to have the machine synthesize its knowledge, rather than use a separate program for each task."

 

Until we can build machines that learn, we will not be able to construct agents with the skills to recognize speech, translate languages, and transact business on our behalf. I am bullish on the prospects for building better and better tools, but I am bearish on the outlook for building effective agents.

 

If the challenge is to synthesize a large diamond, then we can get from here to there using tools. The bets that I would make against Kurzweil are where his predictions presume that we can develop agents.

 

Arnold Kling is the author of Learning Economics

 

Categories:
|

TCS Daily Archives