Kevin wrote:
!!!!!!!!!!!!!!!
*************
In practice, it seems that an AGI is likely to have
an "owner" or a handful of them, who will have the kind of power you
describe. For instance, if my team should succeed in creating a true
Novamente AGI, then even if others participate in teaching the system, we
will have overriding power to make the changes we want. This goes
along with the fact that artificial minds are not initially going to be
given any "legal rights" in our society (whereas children have some legal
rights, though not as many as adults).
************
Would this overriding occur because the person
carries more weight with Novamente, or would they need to go in and altar
the structure\links\nodes directly to affect the
change?
!!!!!!!!!!!!!
Either case could occur. In Novamente, it is possible to assign
default "confidence levels" to information sources, so one could actually
tell the system to assign more confidence to information from certain
individuals. However, there is a lot of flexibility in the design,
so the system could definitely evolve into a configuration where it worked
around these default confidence levels and decided NOT to assign more
confidence to what its teachers told it.
"Going in and altering the structure/links/nodes directly" isn't
always difficult, it may just mean loading a script containing some new
(or reweighted) nodes and links.
!!!!!!!!!!!
*********************************
At least two
questions come up then, right?
1) Depending on the AGI architecture, enforcing one's opinion on the
AGI may be very easy or very difficult. [In Novamente, I guess it
will be "moderately difficult"]
***********************************
That's the crux of the matter
isn't it? Wouldn't it be easy to enforce an opinion while Novamente
is in its formative stages, versus when a large foundation of knowledge is
in place?
!!!!!!!!!!
Yes, that's correct.
**********************************
!!!!!!
Suppose I am overtaken by greed, and I happen to get my
hands on a baby Novamente. I teach it that it should listen to me
above others. I also teach it that it is very desirable for me to
have alot of money. Novamente begins to form goal nodes geared
towards fulfilling my desire for wealth. I direct it to spread
itself on the internet, and determine ways to make me money, preferably
without detection. Perhaps it could manipulate markets, I don't
know. Or perhaps it could crack into electronic accounts and
transfer the money to yours truly.
What's to stop\prevent this? In a real sci fi scenario, perhaps
for your next book, could we have NOvamentes "fighting" Novamente's?
!!!!!!!!
There is nothing in the Novamente architecture preventing this kind
of unfortunate occurence. This has to do with the particular system
of goals, beliefs and habits inside a given Novamente system, rather than
with the AI architecture itself.
!!!!!
This all goes to my concern regarding morality. I know you
resist the idea of hard coding morality into the Novamentes for various
reasons. Perhaps as an alternative, the first Novamente could be
trained over a period of time with a strong basis of moral rules(not
encoded, but trained). Then any new Novamentes would be trained by
that Novamente before being released to the public domain, making it
nearly impossible for the new Novamentes to be taught
otherwise.
!!!!!!
This is something close to what we have planned.
Several others have asked me about this, and I have promised to write
a systematic (probably brief) document on Novamente Friendliness sometime
in early 2003, shortly after finishing my work on the current draft of the
Novamente book.
!!!
I know some of this stuff is a bit out there, but
shouldn't we be considering this stuff now instead of later??
!!!
It definitely needs to be thought about very hard before Novamente
reaches chimp-level intelligence. And in fact I *have* thought about
it pretty hard, though I haven't written up my thoughts much (as I've
prioritized writing up the actual design, which is taking longer than I'd
hoped as it's so damn big...).
Right now Novamente is just a software core plus a bunch of
modules-being-tested-but-not-yet-integrated, running on top of the
core. So we have a whole bunch of coding and (mostly) testing and
tuning to do before we have a system with animal-level intelligence.
Admittedly, though, if our design is right, the transition from
animal-level to human-level intelligence will be a matter of getting more
machines and doing more parameter-tuning, it won't require introduction of
significant new code or ideas.
Having said that I've thought about and will write about it, however,
I have a big caveat...
My strong feeling is that any theorizing we do about AI morality in
advance, is probably going to go out the window once we have a chimp-level
AGI to experiment with. the important thing is that we go into that
phase of experimentation with the right attitude -- with a realization
that training the system for morality is as important as training it for
intelligence -- and with a careful approach.
-- Ben
