David Hart wrote:
On Sun, Oct 5, 2008 at 3:55 PM, Brad Paulsen <[EMAIL PROTECTED]
<mailto:[EMAIL PROTECTED]>> wrote:
More generally, as long as AGI designers and developers insist on
simulating human intelligence, they will have to deal with the
AI-complete
problem of natural language understanding. Looking for new
approaches to this problem, many researches (including prominent
members of this list) have turned to "embodiment" (or "virtual
embodiment") for help. IMHO, this is not a sound tactic because
human-like embodiment is, itself, probably an AI-complete problem.
Incrementally tackling the AI-complete nature of the natural language
problem is one of the primary reasons for going down the virtual
embodiment path in the first place, to ground the concepts that an AI
learns in non-verbal ways which are similar to (but certainly not
identical to) the ways in which humans and other animals learn (see
Piaget, et al). Whether or not human-like embodiment is an AI-complete
problem (we're betting it's not) is much less clear compared with
whether or not natural language comprehension is an AI-complete problem
(research to date indicates that it is).
My argument is not that natural language understanding should be pursued
one way rather than another. It is that it should NOT be pursued at all
for AGI 1.0. And, especially, not by simulating human-like embodiment.
Of course, if you insist on defining AGI 1.0 as ONLY human-like AGI (i.e.,
AGHI), then NLU becomes pretty much a requirement. Also, then, the
difficulty of this problem makes using embodiment (or virtual embodiment)
seem like a good idea. But, again, the "hidden" assumption is that AGI 1.0
must be AGHI 1.0. IMHO, we don't need embodiment and we don't need NLU in
AGI 1.0. If we build AGI 1.0 correctly (avoid what are, for human-like
intelligences, AI-complete issues), it will be able to help us solve many,
if not all, of the AI-complete problems we currently face. In addition, it
could help us decide if having NLU is worth the effort and if embodiment is
worth the risk. NLU may be important for humans. I doubt AGI 1.0 will
care. Same for embodiment.
I also have problems with "...incrementally tackling the AI-complete nature
of natural language processing." The reason AI-complete problems are
AI-complete problems is that they have historically not fallen to the
incremental approach. I'd feel better if the AGHI folks were attempting to
tackle the NLU problem based on a well-thought-out "Grand Theory." At
least that way, they'd fail quicker and we could get serious about building
non-human AGI sooner.
We should, first, develop a non-human, non-embodied AGI that is designed to
help us break the human-machine NLU barrier. THAT's the kind of incremental
approach I think we should be talking about. If we continue to insist on
AGHI as a first step, we're just going to keep banging our heads against
the same old, fifty-eight-year-old wall.
Insofar as achieving human-like embodiment and human natural
language understanding is possible, it is also a very dangerous
strategy. The
process of understanding human natural language through human-like
embodiment will, of necessity, lead to the AGHI developing a sense
of self.
After all, that's how we humans got ours (except, of course, the
concept preceded the language for it). And look how we turned out.
The development of 'self' in an AI does NOT imply the development of the
same type of ultra-narcissistic self that developed evolutionarily in
humans. The development of something resembling a 'self' in an AI should
be pursued only with careful monitoring, guidance and tuning to prevent
the development of a runaway ultra-narcissistic self.
I don't disagree. But, I think your characterization of the human sense of
self as being "ultra-narcissistic" is just inflammatory rhetoric. Your
statements agreeing with me are interposed here before the paragraph from
my original post below that says, basically, the same thing. I just
wouldn't recommend trying to cut off an AGHI's food (power) supply. It
could very well be the last thing you'd do in your puny little biological
life. Human sense of self has, first, to do with survival, not narcissism.
An AGI with no human-like sense of self would just take a nap.
I realize that an AGHI will not "turn on us" simply because it
understands that we're not (like) it (i.e., just because it acquired
a sense of self).
But, it could. Do we really want to take that chance? Especially
when it's not necessary for human-beneficial AGI (AGI without the
"silent H")?
Embodiment is indeed likely not necessary to reach human-beneficial AGI,
but there's a good line of reasoning that indicates it might be the
shortest path there, managed risks and all. There are also significant
risks to be faced (bio/nano/info) for delaying human-beneficial AGI
(e.g., because of being overly precautious about getting there via
human-like AGI).
As noted above, IMHO what's going to delay development of human-beneficial
AGI is exactly what you claim will shorten the path to same. Unfortunately,
as long as the mainstream AGI community continue to hang on to what should,
by now, be a thoroughly-discredited strategy, we will never (or too late)
achieve human-beneficial AGI.
Brad
-dave
------------------------------------------------------------------------
*agi* | Archives <https://www.listbox.com/member/archive/303/=now>
<https://www.listbox.com/member/archive/rss/303/> | Modify
<https://www.listbox.com/member/?&>
Your Subscription [Powered by Listbox] <http://www.listbox.com>
-------------------------------------------
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69
Powered by Listbox: http://www.listbox.com