Eric Baum wrote:
> I believe these issues (as well as Richard's) are well addressed.
> What is Thought? has whole chapters about how exploiting structure
> is a distinct, hard problem from simply finding compact descriptions
> or representations.

Ok, clearly reasonably compact representations make this easier
by reducing the size of the associated search spaces (and doing a
certain amount of equivalent representation collapse 'for free'), but
devoting additional computing time to squeezing the last order of
magnitude out of the byte-level representation size doesn't seem
worthwhile (particularly if the system's base prior is correctly set
up for the 'convenient' representation format in question, rather
than relying on basic Kolmogorov). I may have read too much into
your argument by assuming that you advocated absolute maximum
compression as a cognitive priority.

Unfortunately my copy of What is Thought is packed away in a box
somewhere at the moment, but I'll try and find the time to review
your arguments there when I unearth it.

> Occam code is posited to exploit structure, not merely be a
> compact description or representation.

Much as I like Occam, that's a pretty strong claim for any
language designed primarily for ease of use by human engineers,
indeed for any 'substrate' without an integral optimisation
mechanism. Many people argued that Lisp and Prolog are/were
'posited to exploit structure' at the language level, and while that
worked well for a few early projects (such as AM) I think the
ultimate error of this assumption was a significant contributing
factor to the AI winter. In our system we're using a language
optimised as much as possible for AI code generation (as
opposed to human use; not an uncommon strategy in AGI),
somewhat like a cross between Erlang, Occam and Flare (the
language the SIAI was trying to design back in 2001). Even still,
almost all the modelling power comes from the code pattern library
and associated mechanisms for reasoning about code, not the
language (which primarily makes the latter easier to write and
provides substrate support for various forms of more sophisticated
reflective reasoning).

Michael Wilson,
Director of Research and Development, Bitphase AI Ltd
Web demos page: http://www.bitphase.com/apex

-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Reply via email to