From: Mitchell Porter (email@example.com)
Date: Tue Jan 28 2003 - 23:19:58 MST
As I see it, SIAI's strategy for directly producing a
Friendly Singularity (and not just increasing its likelihood
through Friendliness evangelism) requires two steps:
1. Learn to make LOGI-like AIs (AIs with the five-level
structure outlined in "Levels of General Intelligence").
2. Equip one such AI with the skill set and value system
needed to be a Friendly seed AI.
Now, just how hard is step 1? Or, to put it differently,
what's the simplest program of which Eliezer might say,
"Yup, that's an example of what I was talking about"?
I think we are in dire need of such didactic prototypes
if there is ever to be a *coding community* working on
this thing. There has to be a common understanding of
the basic concepts, and I'm not convinced that that exists
So - there are five levels: code, modality, concept,
thought, deliberation. What I'm going to do is just
informally sketch something which *sounds to me* like
it has all those levels.
Code - no problem here, every program is made of code.
Modality - Above all, a modality seems to be a *feature
extractor*, operating in a specific domain. The input is
some sort of raw data set, static or dynamic, and the
output is a representation as a list of features, or
objects with features. Well, neural nets can do all of
Concept, thought, deliberation - In LOGI, these are
described as analogous to word, sentence, stream of
consciousness. So it seems that *propositional content*
(think Prolog) might be enough for those top two
levels - 'sentences' in some internal language of
representation. The AI needs to produce descriptive
sentences which express what its modalities tell it,
and normative sentences which express its intentions.
If you can get those, there is a wealth of symbolic AI
work on putting them to use in a rational fashion.
So, the key concept seems to be 'concept', and the way
it bridges the gap between subsymbolic feature extraction
and symbolic-level propositional processing. Well, one
simple way to do that is to have concepts which are just
lists of features. Each modality outputs an inventory
of features possessed by its current input; the art of
appropriate concept formation is all about picking out
just a few of those features as 'co-relevant', worthy
of being grouped together. And Copycat provides a model
for *that* process.
It looks to me like we already have all the ingredients,
not for a seed AI, but certainly for a "general AI" as
defined by LOGI.
Add photos to your messages with MSN 8. Get 2 months FREE*.
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT