Re: Metarationality (was: JOIN: Alden Streeter)

From: Eliezer S. Yudkowsky (
Date: Sat Aug 24 2002 - 23:23:43 MDT

Ben Goertzel wrote:
>> Everything that works is a form of rationality; works because it is
>> rational; and is rational because it works.
>> To be precise, everything that works noncoincidentally, with a
>> probability greater than sheer random chance would predict, is a form
>> of rationality; works because it is rational; and is rational because
>> it works.
> Well, that is one definition of the term "rationality."
> However, it's a very general definition.... It certainly does not
> match the standard dictionary definition, and I think it doesn't match
> the common-language definition.

Categories are a form of intelligence; they are a way of slicing up the
universe such that things which can be perceptually placed in a category
can be assumed to have other properties which are not perceptually
apparent. Categories can be useful or nonuseful according to whether they
slice up the universe along the right joints.

Common-language categories related to cognition and intelligence are often
almost completely unusable.

> If we accept your definition, we then need another word for the process
> of "conscious rational thought", which is only one kind of
> "rationality" according to your definition.

The word I use for this is "deliberation".

>> The visual cortex is a form of rationality.
> If so, it's a badly flawed form, due to errors like I mentioned
> (misestimating the distance to faraway objects in the desert...).

Yes, everything can be seen as a flawed or limited form of rationality.

>> That's what the Bayesian Probability Theorem *is* - a *universal*
>> description of the way in which things can be evidence about other
>> things.
> This seems to me to be a large overstatement.
> Bayes' Theorem is a powerful tool (and one of the central parts of the
> Novamente reasoning module).

Explicit, abstract use of Bayes' Theorem, seen deliberatively as Bayes'
Theorem, is a tiny subset of the universal domain governed by the BPT.

> However, like the rest of elementary probability theory, it pertains to
> a set of probabilities defined over a fixed "universal set."
> Defining the universal set is a serious issue in any practical
> situation. You can say "define it as the set of all entities ever
> observed by the mind doing the reasoning". But this doesn't really
> work, because we *hear* about entities via linguistic communication,
> including many entities we haven't seen. I may want to include Beijing
> in my universal set for my internal probabilistic inference, even
> though I have never been there or seen it.

You must learn to see the BPT flowing underneath the surface of all
cognition, like blood beneath skin.

In this case, you have encountered a set of sensory experiences which lead
you to believe that Beijing exists - other people producing reports about
Beijing. Since you expect that people will report that Beijing exists in
cases where Beijing exists, and that they will not report that Beijing
exists if it does not exist, this is, under the BPT, evidence; it is a
case of a thing being evidence about another thing.

In terms of the BPT, this is no different from seeing photons reflecting
from Beijing in person, or from accessing the deposited memories
afterward. These forms of evidence differ in strength and detail, but are
all equally governed by the BPT.

> So then you can't rationally posit a mind that operates only using
> elementary probability theory. You must have two processes in a mind,
> at least:
> -- probabilistic inference (implemented in one of a huge number of
> possible ways) -- a process for defining the universal set (either once
> and for all, or contextually)
> But then how does this second process take place? If you say it takes
> place by prob. inference you get a regress. But if not, you must posit
> some other process, and you must admit that the conclusions of the
> prob. inference (logical reasoning) will be relative to the operation
> of this other process, rather than "objectively rational" (if
> rationality, to you, is defined by the laws of probability theory).

Not at all. You must learn to see the BPT flowing underneath the surface
of all reality.

Reasoning using sensory information takes place against the universal
priors of the brain (the regularities that will form given an exposure to
a given set of experiences). The brain is constructed by evolution, which
can be seen as a form of induction on the universal prior of the set of
possible DNA strings. DNA was arrived at from the original
self-replicating molecule, which was selected from the universal priors
defined by the set of possible molecules. This set of universal priors in
turn arises from the laws of physics. If you ask me which set of priors
our laws of physics were selected from, I don't know. But if you're
trying to be intelligent in our particular universe you might as well
start with the physical "priors" defined by our laws of physics, and in
fact, you will.

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

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:40 MDT