OPINION - On Cyc and EURISKO

The fundamental problem with symbolic AI is that logic is bounded. There is a shape to any corpus of logical expressions that can be programmatically generated via following a set of manipulation rules. And due to the rigid manner in which the corpus evolves, the corpus is bound by its initial position and the rules of manipulation. This is a problem when the corpus has to choose the direction in which it wants to evolve itself further, because the corpus itself is a product of following a set of rigid manipulation rules. Any heuristic rule that can be devised is often not self-referential, and even if it is, it has a limited scope of self-inspection. This is because the heuristic rule itself must be a product of the same mechanism which generated the corpus of logical expressions: the rules of manipulation. And unfortunately, the rules of manipulation are often atomic, often akin to the fundamental laws of logic (e.g. modus ponens).

I think it is often hard for people to imagine the scenario that a more effective heuristic algorithm would be deeply rooted in the laws of probability. The strength of this approach is that if one were to develop a heuristic algorithm like this, then different components could evolve in a synchronized manner purely through emergent behavior. A single change could affect a large number of components not because it has been coded to be that way, but because the emergence of structure is left to the laws of chaos. Whatever structure has stabilized represents a comprehensive integration of everything that has been witnessed. And to change one part of it also in some sense changes the whole of it. There would be no isolation between information, no grammatical structure that designates a beginning or an end. Whereas in the symbolic AI universe each expression can be interpreted in isolation (and in some sense represents a self-contained bundle of information), in non-symbolic AI, the totality of information can be encoded yet none of its individual parts can be visible.

I believe Cyc and EURISKO will never achieve AGI. At least not without leveraging statistical methods. I also have the feeling that Cyc and EURISKO are a philosopher's imagination of artificial intelligence. Both have a Wittgenstein flavor to them, in the sense that both attempt to arrive at the meaning of words by analyzing the grammar of expressions (and subsequently the use of the words within the language).