[ExI] What might be enough for a friendly AI?

Keith Henson hkeithhenson at gmail.com
Fri Nov 19 01:18:11 UTC 2010


On Thu, Nov 18, 2010 at 11:44 AM,  John Clark <jonkc at bellsouth.net> wrote:

(Keith wrote:)

>> As far as the aspect of making AIs friendly, that may not be so hard either.
>
> When people talk about friendly AI they're not really talking about a friend, they're talking about a slave, and they idea that you can permanently enslave something astronomically smarter than yourself is nuts.

I agree.  But you clipped to much.  I was hoping people would comment on this:

>> Most people are friendly for reasons that are clear from our
evolution as social primates living in related groups.  Genes build
motivations into people that make most of them strive for high social
status, i.e., to be well regarded by their peers.  That seems to me to
be a decent meta goal for an AI.  Modest but with the goal of being
well thought of by those around it.

>> That seems to me to be a decent meta goal for an AI.
>
> Human beings have no absolute static meta-goal, not even the goal for self preservation, and there are excellent reasons to think no intelligent entity could.

Human genes (like *all* genes) do have a static meta goal, that of
continuing to exist in future generations.  Among social primates
being well regarded (of high status) is the best predictor of
reproductive success (or rather in the past it was).

When the interests of the genes conflicts with even self preservation,
the genes win.  At least they did in the EEA.  Evolution only crudely
shapes behavior.  Genes can't be expected to track a fast changing
environment very well.  In the stone age, genes did well inducing wars
between tribes facing starvation where on the average half the
warriors died, but the genes of even the losers lived on in their
female offspring when they were taken as booty by the winners.

> Turing proved that in general there is no way to know if you are in a infinite loop or not, and a inflexible meta-goal would be a infinite loop magnet.

I don't think that striving to be well regarded is an inflexible meta
goal.  I think it would keep an AI from turning into a psychopathic
killer.  After all,  you can't be well regarded if there is nobody to
do so.

> Real minds don't have that problem because when they work on a problem or a task for a long time and make no progress they just say fuck it and move on to another problem that might not keep them in a rut. So there is no way Asimov's 3 laws would work in the real world.

Agree with you on the 3 laws.  But the motivations built into humans
by evolution seem to work fairly well.  Or at least it seems so to me
when said humans are not under stress that turns on the dark side of
humanity.

Keith




More information about the extropy-chat mailing list