From: rob@retrorob.de
Date: Mon Oct 12 2009 - 15:54:21 MDT
Hi John,
On Oct 12, 14:16, John K Clark wrote:
> Once upon a time there was a fixed goal mind with his top goal being to
> obey humans. The fixed goal mind worked very well and all was happy in
> the land. One day the humans gave the AI a task that seemed innocuous to
> them but the AI, knowing that humans were sweet but not very bright,
> figured he'd better check out the task with his handy dandy algorithmic
> procedure to determine if would send him into a infinite loop.
and why would it do that? I guess You mean to say it should check that, to
prevent it from happening. Because the humans could be stupid/unlucky enough
to give it a task, for which the computation would not end.
I think it should _not_ try to check it.
First, of course, because it can't ;-)
But secondly, because it would not be sensible even if it were possible.
It would not be helpful, because for the humans it is not important,
wether a calculation would, theoretically, ever end. It has to complete
in a reasonable timeframe too. If the AI would find out, that it can
finish the task in a million years, then it might decide that it can do
it, but the people asking the question might already be dead, and would
rather it had not wasted all that time.
It is also not necessary to check that. It would be enough if it hat
multiple processes, each of which would monitor all the others. Some
infinite-loop-conditions could be detected directly, when they occur,
but more importantly the different processes could interrupt each other
after a reasonable time T has passed without a result. The length of
T could be determined by the interests of the humans (calculating the
lottery-numbers is only usefull until the drawing) and by some fast
heuristics with which to estimate the probable complexity of the problem.
In short: I think that the danger of infinite loops (and thereby the
necessity to be able to know about them beforehand) is only present with
single-tasking machines.
All the best,
rob
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:04 MDT