Rationality, intelligence, evolution

From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Wed Sep 18 2002 - 09:58:47 MDT

Gordon Worley wrote:
> Yes, vastly different. Different enough that I would avoid using the
> term `intelligence' when discussing evolution to avoid confusion. It's
> useful only for showing that all processes that draw upon rationality
> exhibit a certain trait which can be thought of as an intelligence
> (goal-oriented work).

Hm. I have to say this sounds a bit confused even to me.

"Work" or "Nonaccidental success": Nonaccidental truthfinding within an
intelligence, defined as the nonaccidental correlation of an internal
pattern with an aspect of reality; and nonaccidental achievement within
any physical process, defined as nonaccidental development to a goal state
that exhibits some degree of empirical regularity across multiple
*classes* of physical systems. That is, the evolutionary "goal" of
successful reproduction appears with some empirical regularity across
widely different complex systems, and motivates much of the substance of
that complexity, which is not something you can say about stellar dynamics.

"Intelligence": A class of physical systems which generates nonaccidental
achievements using an internal model of aspects of reality.

"Rational physical process": A physical process which can be viewed as an
imperfect approximation of the Bayesian Probability Theorem; in that, with
some degree of empirical regularity, the physical process encounters (some
classes of) variables, and undergoes a perseverant change depending on the
encountered value of that variable, in such a way that it sometimes
enables future achievements (as defined above) because of empirical
regularities within the value of that variable or empirical regularities
connecting that variable to other variables. Since the Bayesian
Probability Theorem describes a normative way for an intelligence to
adjust the expected value of future variables using a previously
encountered variable as evidence, rational physical processes such as
evolution can be viewed as working because they reflect, however
imperfectly and partially, the Bayesian Probability Theorem. In
particular, the strategies employed by these physical processes are
fundamentally within the purview of rationality, in that a purely rational
Bayesian reasoner can understand how they work and could choose to employ
them if the computational and physical capacity to do so existed - there
is no mystical higher force at work.

"Evolution": An accidental, imperfect approximation of the Bayesian
Probability Theorem, which, insofar as it produces nonaccidental
successes, can be viewed as producing them because of a physical process
which reflects the Bayesian Probability Theorem; changes to the genetic
distribution of a future population are based on the environmental
successes of organisms in the past population, but environmental success
correlates to some degree with differential DNA strings, and differential
DNA strings achieve differential successes with some degree of empirical
regularity. This physical process can be viewed as a specialized,
imperfect version of the Bayesian Probability Theorem which reasons about
the likely success of future organisms using the likely success of past
organisms - not really a *flawed* version of the BPT, however; rather a
physical process which, insofar as it works, works because it imperfectly
mirrors the BPT.

"Humans": Nonaccidentally evolved imperfect approximations of the
Bayesian Probability Theorem; intelligent systems which, insofar as they
work, work because the physical process of those intelligent systems
imperfectly mirrors the BPT with respect to some task or subtask. Gordon
is, as far as I know, correct in stating that all known nonaccidental
successes have evolution somewhere in their causal chain; evolution,
humans, and human technology.

"Rationality": I think this ends up referring to the degree which human
reasoning, especially high-level deliberation about declarative beliefs,
mirrors the BPT consistently and correctly, rather than mirroring it in
occasional flashes which get small amounts of work done. Also the degree
to which truthfinding in particular functions consistently and correctly,
perhaps in despite of other evolutionary adaptations which produce
nonaccidental achievement in reproduction at the expense of nonaccidental
truthfinding. Rationality as a discipline exists because of those
empirical regularities in the degree of human rationality which depend on
variables subject to human volition (such as philosophical beliefs).

Eliezer S. Yudkowsky http://intelligence.org/
Research Fellow, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT