What Prevents Seed AI?

[Luminal]

Just wondering... I don't understand why a true 'seed' AI program has never been developed after numerous decades of this concept existing in some form or another.

 

In it's most simple form, seed AI does this:

 

1) A program with preexisting functionality performs some change upon itself; any part of the program can be. The change can be entirely random or partially guided by the programmer.

 

2) If the program crashes or cannot compile, return to step 1. Otherwise, run the program again and test its performance against its original performance.

 

3) If new performance is superior in its functionality, then the new version is used and returned to step 1. If inferior, return to step 1 with original version.

 

Eventually, tests for different functionality can be implemented, whether that be pattern recognition, language, or other intelligent behavior. The concept being, if implemented correctly, that there will be an exponential feedback of positive changes after enough time. Thus my question: why hasn't something like this ever moved past the initial stages and entered an exponential feedback cycle?

 

In essence, this is similar to natural evolution, except that it has several significant advantages, namely: the ability to store previous versions and step back to them when insurmountable obstacles are encountered; the synergistic relationship of existing intelligent organisms (humans) helping to guide their evolution; generations being as fast as the execution of the program (often measured in seconds), whereas in nature the shortest generations are ~20 minutes and the longest generations are decades or even centuries.

Edited November 12, 2009 by Luminal

[Kyrisch]

The main problem with this is that programming a way for the program to determine its own fitness is incredibly difficult. How would one go about programming a function which calculates fitness that's general enough to recognize all forms of 'higher' complexity as the ones you listed above?

[bascule]

Just wondering... I don't understand why a true 'seed' AI program has never been developed after numerous decades of this concept existing in some form or another.

 

Well, before you can talk about seed AI (recursively self-improving AI) we have to get to strong AI first. We aren't anywhere close to a computer as intelligent as your average 4 year old child.

 

One of the biggest problems is that most of the research has been misguided. Rather than trying to crib from biology, researchers attempted to invent their own ideas about how the mind works out of thin air and implement those.

 

Perhaps the book that makes this clearest is Marvin Minsky's The Society of Mind. Now don't get me wrong, it's a good book and I do think he gets a lot right, but at the same time this entire hypothesis is untestable and cannot be accurately tested to determine if it's how our minds work. Worse, the book has its own nomenclature for describing everything.

 

The other problem is one of the sheer amount of computation involved. Our brains contain a hundred billion neurons, each of which has the equivalent complexity of an integrated circuit. Right now computers simply aren't fast enough to perform all the calculations necessary to rival a human brain.

 

Things are moving in the right direction though. Software is being developed that is designed to perform similar functions to the human neocortex. This software is already being used to develop intelligent narrow AI products, and, over time, could form the basis of a strong AI system.