justin corwin wrote:
It is entirely possible to build an AI in such a way that the general
course of its behavior is as reliable as the behavior of an Ideal Gas:
can't predict the position and momentum of all its particles, but you
sure can predict such overall characteristics as temperature, pressure
and volume.
This is the only claim in this message I have any disagreement with
(which must be some sort of record given my poor history with
Richard). I agree that its true in principle that AIs can be made this
way, but I'm not yet convinced that it's possible in practice.
[Heck, yeah: I had to check to see if maybe you were some other justin
corwin ;-)]
I agree with your caution about that claim.
For what its worth, though, what I had in mind was a motivational system
that acts like a point moving around in an unconstrained way in a
multidimensional space (hence my thermodynamics metaphor), BUT with the
shape of that space heavily distorted so that it has one enormously deep
basin of attraction. To be stable and safe, it must stay in the basin.
Can it get out? Yes, in principle. Mean time between escapes from
the basin? Greater than lifetime of the universe by a huge factor. Is
it therefore guaranteed to be safe? No. Is it ever likely to get out?
Not a chance.
(I haven't said how to do this. What I just gave was a summary of
overall dynamics of a particular system).
In the same way, all the molecules of an ideal gas could, in theory,
just happen to divide into two equal chunks and head toward opposite
ends of the box at full speed. They have the freedom to do so. We just
don't have to worry about it happening, for obvious reasons.
The beauty of the approach is that the system can know perfectly well
that it has been engineered that way by us. But it does not care: if
it starts out being empathic to humanity, it will not deliberately or
accidentally change itself to contradict that initial goal. It will
have plenty of freedom to do so in principle, but in practice it will
know that the consequences could be disastrous.
I realise that I have been tempted to explain an idea in partial,
cryptic terms (laying myself open to requests for more detail, or
scorn), so apologies if the above seems opaque. More when I get the time.
Richard Loosemore.
It may be that the goals of and motivations from such artificial
systems are not one of those characteristics that lies on the surface
of such boiling complexity, but within it. I have the same
disagreement with Eliezer, about the certainty he places on the future
characteristics of AIs, given that no one here is describing the
behavior of a specific AI system, such conclusions strike me as
premature, but perhaps not unwarrented.
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/[EMAIL PROTECTED]
