Re: Metarationality (was: JOIN: Alden Streeter)

From: Eliezer S. Yudkowsky (
Date: Sun Aug 25 2002 - 23:55:19 MDT

Ben Goertzel wrote:
> Eliezer, you're using the word "rational" to mean "anything that works".

Actually, I'm using it to mean "Anything knowable to a Bayesian reasoner"
and am then making the nontautological statement that "everything that
works noncoincidentally" is equivalent to "everything knowable to a
Bayesian reasoner".

> So if you have an organism with a goal, then any process that it uses that
> significantly helps it to achieve that goal is "rational" by your
> definition. (Correct me if I've paraphrased your idea incorrectly, please.)
> This is a meaningful sense of the word "rational", but it's a very broad
> one.
> In this broad sense, it's true, even processes that have nothing to do with
> *reasoning* as commonly conceived may be viewed as "rational."
> For instance, a bird, acting purely on instinct, is acting rationally when
> it mates with another bird, because it's acting in a way that "works" in
> terms of achieving its goal of reproducing itself.

In what way does this have nothing to do with reasoning? It is a product
of evolution's version of induction. Furthermore, the fact that acting in
this way will increase the probability of reproduction is quite knowable
to a Bayesian reasoner. A perfect Bayesian reasoner might use a more
optimal method of reproduction, but an imperfect Bayesian reasoner might
end up employing exactly this method. Because evolution *is* that kind of
imperfect reasoner.

> I generally have been using the word "rational" to mean something narrower.
> Basically, to mean "following a process of logical inference", meaning a
> process inside some mind in which premises are represented, conclusions are
> represented, and processes for getting from the premises to the conclusions
> are carried out. Note that this sense of "following a process of logical
> inference" is implementation-independent; it applies to neural nets and so
> forth as easily as to minds embodying some sort of explicit
> symbol-manipulation component.

Yes. And because you use "rational" to denote only this kind of thought,
you can get away with telling yourself that "both 'rationality' and
'irrationality' are necessary to thought", thereby avoiding the necessity
of getting rid of comforting irrational thoughts.

Which it is why it's so important to realize that verbal thought is verbal
thought, intuition is intuition, and that both can be either "rational" or
"irrational". It's very easy to find a good name, other than
"rationality", for verbal reasoning or deliberative reasoning or conscious
reasoning... in fact, I tossed out three right there. But I don't know
another good name, besides "rationality", for the quality which correct
intuitions and correct declarative thought share with each other, and do
*not* share with arational sources of perceived support such as "that's a
comforting thought".

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