Accelerating Future Transhumanism, AI, nanotech, the Singularity, and extinction risk.

Ray Solomonoff on Speed of AI Takeoff

In 1985, Ray Solomonoff offered his thoughts on six milestones in AI and the economic and technological growth that might be expected when generally intelligent AI is developed. The paper is called "The Time Scale of Artificial Intelligence: Reflections on Social Effects".

Here is the abstract:

Six future milestones in AI are discussed. These range from the development of a very general theory of problem solving to the creation of machines with capacities well beyond those of a single human. Estimates are made for when these milestones will occur, followed by some suggestions for the more effective utilization of the extremely rapid technological growth that is expected.

When I read lines like that last sentence, what I see nowadays is "extremely scary technological growth". Rapid growth is scary when that growth is controlled by systems that may not optimize reality in ways that we explicitly value. (See "The Future of Human Evolution" for an explanation.)

A select milestone:

Milestone C. A critical point in AI development would be a machine that could usefully work on the problem of self-improvement. Newell and Simon were not successful in their attempts to get their "General Problem Solver" to improve it's own methods of operation. While Lenat's "Eurisko" has been successful in several problem areas, he has not been able to get it to devise good heuristics for itself. He is, however, optimistic about the progress that has been made and is continuing this work.

Eurisko eventually led to the creation of Cyc, which appears to be of limited use.

It should be noted that AI "self-improvement" should be viewed as a special case of an AI's general talents for understanding an object, evaluating its purpose, and improving it with respect to that purpose. (Sometimes people make unwarranted distinctions between an AI modifying itself and modifying the world.)

How about some more milestones:

Milestone D. Another milestone will be a computer that can read almost any English text and incorporate most of the material into its data base just as a human does. It would have to store the information in a form that is useful for solving whatever kinds of problems it is normally given.
Since there is an enormous amount of information available in electronic data bases all over the world, a machine with useful access to this information could grow very rapidly in its ability to solve problems and in a real sense in its understanding of the world.

Milestone E will be a machine that has a general problem solving capacity near that of a human, in the areas for which it has been designed -- presumably in mathematics, science and industrial applications.

Milestone F will be a machine with a capacity near that of the computer science community.

Milestone G will be a machine with a capacity many times that of the computer science community.

Here's another bit from later on, analyzing the potential impact of Milestone G:

The last 100 years have seen the introduction of special and general relativity, automobiles, airplanes, quantum mechanics, large rockets and space travel, fission power, fusion bombs, lasers, and large digital computers. Any one of these might take a person years to appreciate and understand. Suppose that they had all been presented to mankind in a single year! This is the magnitude of "future shock" that we can expect from our AI expanded scientific community.

Scanning over the paper, it still seems like Solomonoff is thinking of AIs as tools or narrow scientists, rather than general agents with the full range of activity that humans have or beyond. In the end, Solomonoff seems to imply that one of the primary benefits of AI will be to allow us to predict and evaluate the future more effectively. But he points out that we will still have to make ethical choices.

H/t to Shane Legg for writing about the paper.