You seem to have forgotten Ben that “magic sauce” is a term used by many
AGI-ers for the “missing but crucial ingredient of AGI* wh. an AGI
projectbuilder claims he will produce at a later stage from his present outline
but never does.
You’ve also forgotten how many times you have indeed promised a missing but
crucial ingredient of AGI would be produced at a later stage with some
publication of yours – and never have produced it.
You have done a QED by now indicating another magic ingredient of AGI, this
time a pattern recognition system . you have never produced any actual
examples of how patterns are relevant to AGI problems, and never will – just
waffled generally. This is complete (if fairly widely-held) nonsense –
By definition, an essential requirement of AGI is to solve problems about
actions and environments that do NOT fit existing patterns. By definition an
AGI must acquire new skills and undertake new kinds of actions – something that
presently defies all narrow AI programs. This is a problematic business –
precisely because the new skill/action is NOT like already known actions – does
NOT fit any existing, known pattern. Mastering a new skill like say the actions
of table tennis after tennis, or vice versa, presents difficulties precisely
because the arm and body actions do not fit the same patterns.
Mastering/understanding physics after chemistry is difficult precisely because
the laws of one do not easily fit the laws/patterns of the other. And so it is
with every subject area in organized knowledge.
Being able to have a conversation with one person is difficult precisely he
does NOT fit the conversational patterns of others. Every one is different.
Talking to Jim about AGI is different from talking to Matt, or Pei, or Aaron,
etc etc – because each one has a different approach to AGI, each one is
individual, and each one’s idiosyncrasies have to be gradually identified in
order to talk to them. They may share some common elements, but overall they
are very different.
Nor do the conversations or posts of any individual taken altogether fit a
distinct pattern. We may have distinctive “styles” of conversation, but those
styles are fluid schemas at best, and nothing like the precise patterns you are
talking about – and result in very diverse, multiform posts. Check your own
posts in this or any other thread. I defy you to identify overall patterns.
I know that you have never actually applied your pattern theories to actual AGI
problems, just as it was clear from your book on creativity that you had never
applied your creativity theories to any actual creative problems that you had
independently researched.
You work by adapting other people’s theories – and unfortunately for you, none
of them apply to building a real AGI., esp. patterns and pattern recognition.
P.S. Some indication of how AGI’s actually adapt to the new is given in our
talking of “a period of adjustment” being required, of “getting the hang” of
things, and “finding our feet” after stumbling and groping around. All of this
adapting to the new has nothing to do with pattern recognition.
A GOS must be designed to enable a robot to endlessly *develop* its actions –
endlessly move along *new* lines - not fit its actions to the same old
patterns and lines. That’s narrow AI.
From: Ben Goertzel
Sent: Thursday, December 27, 2012 10:56 PM
To: AGI
Subject: Re: [agi] The Only Test of AGI
Neither Ben nor anyone else in AGI is directly addressing the problem of a
take-off system – or indeed has a clue – wh. is why you can immediately write
off Opencog and other such efforts. They have absolutely nothing to do with
AGI/ take-off – wh. is also why Ben et al have always resisted any form of test
– they always have and always will fail any test of take-off/generality. (It’s
not just me BTW – many have remarked that Ben et al’s “magic sauce” is not
there – not even the idea of one)
The idea that a "magic sauce" is needed for AGI is a mystical delusion,
redolent of vitalism in biology...
The statement that I resist any form of test is a bald-faced lie. I don't
think that testing a completed human-level AGI is a particularly hard problem,
and I think it's a useful thing to do. The Turing Test is an OK one (if it does
on for an hour or more), or the test of having a robot pass the third grade,
etc. etc. I am skeptical of quantitative metrics for early-stage partial
progress toward human-level AGI, because I haven't yet seen any that aren't
either
-- requiring a system already 80% of the way to human-level AGI
-- too easily game-able by narrow-AI systems written especially to pass the test
...
The magic of general intelligence is simply this: A pattern recognition system
that can recognize patterns in its environment and *itself*, including patterns
regarding which actions tend to achieve which objectives in which contexts.
The challenge of general intelligence is: Recognizing a sufficient scope of
patterns, within a relevant and broad set of contexts, within the limited
compute resources available....
Meeting this challenge seems, so far as I can tell, to require a fairly complex
and multifaceted software system with interdependent parts; which makes
building AGI a major engineering and algorithmic challenge.
That's not as romantic as daydreaming about some "magic sauce" that you can
just pour into your robot's head to make its wiring or its software get smart
-- but it's the reality...
-- Ben G
AGI | Archives | Modify Your Subscription
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-f452e424
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-58d57657
Powered by Listbox: http://www.listbox.com
