[p2p-research] is the mind a computer

[J. Andrew Rogers]

On Thu, Nov 12, 2009 at 9:01 AM, Ryan Lanham <rlanham1963 at gmail.com> wrote:
> I've just become to read it because skimming proved useful.  I'm also
> looking at his web and blog...both informative.  He certainly doesn't much
> cite mainstream AI people much...almost like this is the rise of a new
> thread.  There must be a story behind where these guys have come
> from...these things usually don't materialize out of the blue.


Legg started working on AI at universities in the 1990s, but largely
discarded that classical AI work as far as I know. He was one of the
very first people to work on general AI using algorithmic information
theory. The problem with virtually all mainstream AI (specifically,
mainstream AI not derived from algorithmic information theory) is that
it is baseless in that there is no strong theoretical basis why any of
it should work. Much of the prior AI research can be derived and
analyzed in the context of current algorithmic information theory, so
there are not many cases where it is meaningful to cite prior research
in AI per se.

Around the 2000-2002 timeframe, a few different people noticed that
you could derive a formal model of universal intelligence from some
concepts in information theory that had been largely ignored.  Marcus
Hutter, mentioned previously, developed the first proper derivation
(google: AIXI) and Shane Legg started working with him. This
theoretical work is provably very powerful and very general, but has
almost no relation to previous AI research.

The research priors for intelligence in computer science are not
previous AI research, but some obscure areas of information theory
such as Solomonoff induction. To the extent prior AI research has some
validity, it can generally be shown to be very narrow cases of the AIT
model.  The AIXI model is quite a bit like the Turing Machine model in
that you can show some deep equivalencies between systems that look
very different on the surface and real-world implementations can only
approximate the general theoretical model.


-- 
J. Andrew Rogers
realityminer.blogspot.com