[ExI] Self improvement was Did Hugo de Garis leave the field?

Mike Dougherty msd001 at gmail.com
Thu Apr 21 07:11:34 UTC 2011


On Thu, Apr 21, 2011 at 1:12 AM, Keith Henson <hkeithhenson at gmail.com> wrote:
> Thought experiment.
>
> Build an AI.  Put a clock rate control in that it can mess with.
>
> If it is measuring it's performance against an outside clock, it is
> going to notice that it does better with the clock rate on fast.
>
> So it sets it's clock rate to fast, effectively becoming smarter.
>
> Seems to me this is self improvement.

I think that's self-evident if we agree on the definition and
measurement of "smarter"

I'd say your own ability to solve any arbitrary engineering problem is
sufficiently beyond my own that even with my local clock running at 2X
you'd still be smarter than me.  So I think the definition of smarter
(especially in light of self-control of clock rate) must include some
normalizing factor such that "smarter" measures useful/effective
output per 'clock.'

I think it would be a good realization for the AI to turn the clock
all the way up - at least to the optimum of physical constraints such
as heat dissipation and energy consumption.  Intentionally consuming
short-term smart-drugs with the long term effect of destroying the
brain is not smart overall.




More information about the extropy-chat mailing list