From: Eliezer S. Yudkowsky (sentience@pobox.com)
Date: Fri May 12 2006 - 09:31:44 MDT
Joshua Fox wrote:
> Eliezer S. Yudkowsky wrote:
>
>> .... the way that some people apply their theories of human
>> irrationality to the larger processes of cognition (in this case,
>> judgments of future scenarios), and come to conclusions as if teh
>> Conjunction Fallacy, and the general lack of logical reasoning skills,
>> were the main determinants in those analyses.
>
> Because flaws in rationality such as the Conjunction Fallacy are not
> explicitly adaptive to the human evolutionary environment, might it be
> that these are necessary side-effects of practical General Intelligence?
> Every design requires compromise in some parameters, and perhaps this is
> an /essential /compromise for the first generation of AGI.
I didn't write that, Loosemore did.
I think some forms of the Conjunction Fallacy will survive in
better-designed minds, others not.
A mind evenhandedly penalizing a surface description by the
improbability of each of its conjuncts does not sound very hard to do.
An AI should not think that Linda is more probably a feminist bank
teller than a bank teller.
Where a possible cause is mathematically simple but very expensive to
find - that is, there's a very simple explanation that retrodicts all
the data, but coming up with that simple explanation is a cognitively
hard task - then a boundedly rational mind may:
1) Be asked "What's the probability of B?", and reply with a low
probability for "B".
2) Then be asked "What about A&B?" (where A is the simple cause of B),
and reply with a higher probability for "A&B?" than was answered for
"B?", but also updating the probability for B alone so that the new
probability given to B is higher than A&B.
-- 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 - 04:00:56 MDT