From: Krekoski Ross (rosskrekoski@gmail.com)
Date: Thu May 08 2008 - 13:23:07 MDT
Actually, now that I think about it-- this precludes a fast take-off with no
end.
Any intelligent system cannot meaningfully increase its own complexity
without a lot of input, or otherwise assimilating complexity and adding it
to its own.
Consider I have an AI of a given complexity, lets call it takeoff(n) where n
is the number of iterations we would like it to run for, conceivably a very
large number. takeoff() itself has a specific complexity.
however, without a large number of inputs, takeoff() cannot increase its own
complexity, since the complexity it reaches at any point in time is
describable with its own function, and a specific integer as an argument.
any AI can therefore really only at most, assimilate the complexity of the
sum of all human knowledge. Without the ability to meaningfully interact
with its environment and effectively assimilate information that is
otherwise external to the sum of human knowledge, it will plateau.
Ross
On Thu, May 8, 2008 at 6:43 PM, Krekoski Ross <rosskrekoski@gmail.com>
wrote:
>
>
> On Thu, May 8, 2008 at 8:59 AM, Stuart Armstrong <
> dragondreaming@googlemail.com> wrote:
>>
>>
>>
>> What makes you claim that? We have little understanding of
>> intelligence; we don't know how easy or hard increases in intelligence
>> will turn out to be; we're not even certain how high the advantages of
>> increased intelligence will turn out to be.
>>
>> It could be a series of increasing returns, and the advantages could
>> be huge - but we really don't know that. "Most likely scenario" is
>> much to strong a thing to say.
>>
>
>
> Yes.
>
> I personally dont have a strong opinion on the probability of either
> scenario just because there are so many unknowns, and we have an effective
> sample size of 1 (ourselves) with which to base all of our understanding of
> intelligence. But I think we should realize one thing--- is it only by
> incredible coincidence that our intelligence is at a level such that we can
> understand the formal properties of our brain, but are just below some
> 'magical' threshold that would allow us to mentally simulate what
> differences in subjective experience and intelligence a slight change in our
> architecture would entail, but just above the threshold where it would be
> possible to do so for 'lower' entities?
>
> I've mentioned this before in various forms but in general I think its a
> fairly under-addressed topic: Can an intelligent system of complexity A,
> perfectly emulate (perform a test run of) an intelligent system of
> complexity A? (for fairly obvious reasons it cannot emulate one of higher
> complexity). It seems possible that an intelligent system of complexity A
> can emulate one of complexity A-K where K is the output of some function
> that describes some proportion of A (we dont know specifically how
> complexity in an intelligent system affects intelligence, except that in a
> perfectly designed machine, an increase in complexity will entail an
> increase in intelligence). I think that because of natural systemic
> overhead, it is impossible for any perfectly designed intelligent system to
> properly model another system of equal complexity. (and indeed no effective
> way to evaluate the model if it could)
>
> This has implications on the rate at which any AI can self-improve-- if K
> is a reasonably significant proportion of A, even a godlike AI would have
> difficulty improving its own intelligence in an efficient and rapid way.
>
> This is also why evolution by random mutation is a slow, but actually quite
> efficient way of increasing intelligence--- we dont want a progressively
> larger, but structurally homogenous system (which actually is not an
> efficient increase in complexity, only size). We want structural diversity
> in an intelligent system, and its not clear how a system can 'invent' novel
> structures that is completely foreign to it. Many of our own advances in
> science, by analogy, arise from mimicry of for example non-human biological
> systems.
>
> Ross
>
>
>
>>
>> Stuart
>>
>
>
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:02 MDT