You may be tired of reading this series about Eliezer Yudkowsky. I know I’m tired of writing it. So let’s try to move towards conclusion.
Computer programmers know that the most important of a programming task is thoroughly understanding the problem and working through the solution. Generally, most of the work of constructing a program occurs before any code is ever written.
So the fact that Yudkowsky has no code examples or testable constructions doesn’t count against him. That’s the final stage, and it would be unreasonable to expect anyone to have reached that level of understanding at this point in AI construction.
But what has he actually accomplished? We’ve previously seen that his hypothesis of ‘Friendly AI’ covers the same ground as Isaac Asimov’s Three Laws of Robotics. The thought experiments Asimov embedded in his fiction anticipated Eliezer’s own thoughts, and Asimov wrote them long before Eliezer was born. What new territory has Eliezer covered?
In my examinations of Yudkowsky’s writings, I can find only two ideas that would seem to qualify. One is his ‘Hostile AI in a Box’ experiment, and the other is the ‘Coherent Extrapolated Volition’ hypothesis.
Hostile AI in a Box is next.