Enlightened OpenCog minds...

71 views
Skip to first unread message

Ben Goertzel

unread,
May 1, 2013, 2:07:07 AM5/1/13
to opencog, hk-a...@opencog.org
Just for general amusement... here are some fluffy speculative
thoughts of mine on why a completed, fully intelligent OpenCog system
might be more "enlightened" than human beings ;-) ...

http://multiverseaccordingtoben.blogspot.hk/2013/05/the-dynamics-of-attachment-and-non.html


... ben

--
Ben Goertzel, PhD
http://goertzel.org

"My humanity is a constant self-overcoming" -- Friedrich Nietzsche

Linas Vepstas

unread,
May 3, 2013, 2:59:51 PM5/3/13
to Ben Goertzel, opencog, hk-a...@opencog.org
Insofar as attachment is like addiction, there has been considerable
progress in understanding the mechanisms of addiction. There was a
Scientific American article on this, maybe 5-10 years ago. The
mechanism was surprisingly complex, and involved the interplay of
maybe 4(?) mechanisms working at the day, week, month and year
timescales. At the one-day scale, you can consciously resist impulses
and urges. Doing so causes changes in certain neurotransmitter
levels, but it took a week of resisting before corresponding changes
occurred in wiring. These in turn altered .. !? ... I don't
remember, but one surprising part was that there was at least one part
of the system which stayed in fixed state for a year or two, before
finally releasing. Failures at the level of any one of these 4(?)
circuits can lead to relapse; and the relapses have a characteristic
signature.

This was used to explain how one can stay clean for 6 months and still
fall off the wagon; it seems like 4-5 years is a more reasonable
time-frame to become completely de-attached from addictive behaviors.

I don't necessarily see that AGI would have a less sophisticated
addictive network, nor that it can avoid these kinds of problems, or
even deal with emotions properly. So for example: the sensation of
danger should lead to a preparation for fight-or-flight, for
self-preservation. For an AGI, this would mean shuttling more power
to motors, shutting down non-essential functions, spending less time
day-dreaming and more time focusing on the immediate situation.
Roughly speaking, this is "an emotion": a deep, fundamental systemic
change with sharp affective changes to go with it. You are not going
to be having a flowery, poetic conversation if you are about to take
off and run.

Once in such a state, it may take a while to wind down from it: put
enough distance between self and danger, for example, but also some
time to resume normal activities. But also normal activities may be
curtailed long enough to take steps to contain the danger: the AGI
could stay single-minded for a while.

Its plausible that returning to the "dangerous place" could re-trigger
the fight-or-flight response, even if the immediate danger has passed.
Various automatic subsystems may still be reacting automatically,
even as the conscious part tries to keep them under control. And
don't tell me that AGI won't have independently-running automatic
subsystems under only partial conscious control. It would be foolish
otherwise: You don't want to trip while walking, just because a
brilliant idea shot into your mind just right now, and used up all
available CPU cycles.

Based in this analogy, I claim that reasonably advanced AGI will feel
something similar to "emotions". One question is whether merely
thinking about a dangerous situation would be enough to provoke an
"emotion". One could probably argue that it should: merely thinking
about going into battle should be enough to empty your bowels, whether
you're a human or a robot.

-- Linas

Ben Goertzel

unread,
May 4, 2013, 12:38:40 AM5/4/13
to linasv...@gmail.com, opencog, hk-a...@opencog.org
Linas,

Depending on the AI's architecture, it might have much more power to explicitly
and deliberately rewire its internal knowledge than humans have...

So when you say

" it took a week of resisting before corresponding changes occurred in wiring."

-- for an appropriately architected AI, desired and known changes in
wiring could
be implemented almost immediately by the AI itself, based on its
high-level goals...

This capability may lead to various pathologies (e.g. wireheading),
but obviously
would make the process of succumbing to or overcoming addictions quite different
for AGIs than for humans...

-- Ben

Linas Vepstas

unread,
May 4, 2013, 2:04:46 AM5/4/13
to Ben Goertzel, opencog, hk-a...@opencog.org
Well, I guess I was trying to say that its not obvious that allowing
immediate drastic changes are a good design principle. I'm tired of
making analogies, but with a little effort, one can find examples
where its good to have inertia, or braking, or damping -- to prevent
oscillations and other pathologies.

The point about addiction is that, for some reason, the human brain
developed this fairly complex, sophisticated mechanism for handling
goal-rewards -- it has both short and long time-scale parts to it.
When it functions poorly, it leads to addiction. But I'm assuming
the complexity is there for a reason, and its not obvious that an
AGI-equivalent goal-reward management system should cut out the three
slow parts, and keep only the fast, immediate part.

--linas

Ben Goertzel

unread,
May 4, 2013, 2:07:44 AM5/4/13
to linasv...@gmail.com, opencog, hk-a...@opencog.org
> The point about addiction is that, for some reason, the human brain
> developed this fairly complex, sophisticated mechanism for handling
> goal-rewards -- it has both short and long time-scale parts to it.
> When it functions poorly, it leads to addiction.

True...

>But I'm assuming
> the complexity is there for a reason,

I guess you have more faith in the wisdom of evolution than I do !!

>and its not obvious that an
> AGI-equivalent goal-reward management system should cut out the three
> slow parts, and keep only the fast, immediate part.

No doubt refining a good AGI goal-reward management system will require lots
of experimentation, and maybe development of some new theories...

I doubt the result will bear much resemblance to the human goal-reward
management system, or will be subject to addiction or rampant irrational
attachment, though ;)

ben

Linas Vepstas

unread,
May 4, 2013, 4:13:29 PM5/4/13
to Ben Goertzel, opencog, hk-a...@opencog.org
On 4 May 2013 01:07, Ben Goertzel <b...@goertzel.org> wrote:

>>and its not obvious that an
>> AGI-equivalent goal-reward management system should cut out the three
>> slow parts, and keep only the fast, immediate part.
>
> No doubt refining a good AGI goal-reward management system will require lots
> of experimentation, and maybe development of some new theories...

How about the idea of maximizing the number of accessible future
possibilities? (aka 'causal entropy')

> I doubt the result will bear much resemblance to the human goal-reward
> management system,

Agreed.

> or will be subject to addiction or rampant irrational
> attachment, though ;)

Not at all obvious. When things fail, they fail spectacularly. I
can't imagine that AGI would be failure-free.

There's some theorem about complexity and debugging: debugging code is
harder than writing it. If you write your code as cleverly as
possible, how will you ever debug it? I think this is true for AGI
as well -- it cannot possibly foresee all the possible outcomes and
failure modes. Nor is it clear that it will have time to fix those it
can foresee...

-- Linas

Ben Goertzel

unread,
May 4, 2013, 11:38:24 PM5/4/13
to linasv...@gmail.com, opencog, hk-a...@opencog.org
>> No doubt refining a good AGI goal-reward management system will require lots
>> of experimentation, and maybe development of some new theories...
>
> How about the idea of maximizing the number of accessible future
> possibilities? (aka 'causal entropy')

Do you mean: Using that as a goal for an explicitly goal-oriented AGI system
(like OpenPsi + PLN, with causal entropy as the sole or primary goal
of OpenPsi)?
In this case, causal entropy would be part of the picture, but would
not dictate the
whole goal-reward management system, right?

Or do you mean: making an artificial life / artificial chemistry type
simulation, in which
complex structures (including potentially intelligences) emerge from a substrate
into which the causal entropy principle is embedded?

-- Ben

Linas Vepstas

unread,
May 5, 2013, 1:22:56 PM5/5/13
to Ben Goertzel, opencog, hk-a...@opencog.org
Well, presumably both. :-)   What's good for the gander is good for the goose.

--linas
Reply all
Reply to author
Forward
0 new messages