Re: Changing the value system of FAI

From: Eliezer S. Yudkowsky (
Date: Wed May 10 2006 - 15:11:11 MDT

Ben Goertzel wrote:
> Well, Godel's Theorem shows that for any reasonably powerful and
> consistent formal system, there are some statements that cannot be
> proved either true or false within that system. Furthermore, many of
> the examples of this kind of undecidable statement happen to be
> "meta-statements" that pertain to the formal system as a whole.
> So, if we have an AI system that operates via consistent application
> of a formal system (e.g. some variant of mathematical logic, including
> probabilistic logic), then there will be some statements about this
> system that cannot be proved true or false within the system.

But the question is what impact this has on *decisionmaking*. What is
the AI prohibited from *doing* as a result of Godel's Theorem? What
changes is it prohibited from making to its own code? Classical
mathematical logic is all about proof, assertion, and believing, rather
than matters of decision theory. Moreover, classical mathematical logic
is about *belief* rather than *anticipation*, in the sense of the
distinction made in _Technical Explanation_ - it isn't about organizing
sensory impressions. We know what happens when a proof system asserts
its own consistency. What if an AI behaves as if a statement that it
proves in Peano Arithmetic has a very low probability of being wrong?
And what happens when the AI rewrites the source code responsible for
behaving as if statements proven in PA have low probabilities of being
wrong? Or rewrites the code that rewrites the code? Does the AI
necessarily have to *believe itself consistent*, in the sense that
causes a formal system to break down, in order for code to rewrite
itself? We know what the Godelian restrictions are - but there's a
difference between knowing that, and being able to say that Godelian
restrictions imply limitations for AIs.

Eliezer S. Yudkowsky                
Research Fellow, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:56 MDT