I do not speak much but follow the discussions of the list. I am currently
an undergraduate student at Lund university in Sweden. At the moment, I'm
primarily still learning and touching different subjects.

I am interested in primarily three things:
* Optimal algorithms for arbitrary well and ill-formed problems. For
ill-formed problems, I do not even know if it makes sense. Well-formed
problems are to some extent solved, which is very exciting. There are some
restrictions with the current algorithms, such as optimal decisions w.r.t.
to limited resources or utility functions contingent the used resources. By
optimal I obviously mean the maximal expected outcome of any policy w.r.t.
to the resource-dependent utility function. I do not only mean the question
of what resources are neede for the maximal expected utility (EU*) without
any resource restrictions for particular problems. The greatest goal of AI,
I do not consider only to achieve human-level intelligence but optimal and
either a formal way to handle ill-formed problems or limited resources would
be a grand next step.
* Proper handling of ill-formed/ill-posed problems. This seems tricky to me.
Many times, statistics may be applied to learn about the problem,
generalize, simplify and solve the simpler problem. Many tests today, mind
you, most of what you discuss on this list, is sampling where one attempts
to learn about the problem at hand. A lot of complications arise with this
and I feel even more lost on how to apporach the first point in this
context.
* The formal computer science foundations of decision science algorithms (
i.e. non-mimicry AI). This class is broader than machine learning and does
include for instance logical inference. A further division would be 1) AI
problems on a wit. Such as logical inference and heuristics that does not
generalize from observations. 2) generalization/machine learning. Here,
heuristic approaches are justified through statistics even if it is not
explicitly in the motivation. The difference from regular AI vs. ML would be
the heuristic classes. A less important goal would also be to formalize all
heuristic approaches with a statistical foundation.

Any comments, criticism or pointers much appreciated.

On 3/26/07, DEREK ZAHN <[EMAIL PROTECTED]> wrote:

David Clark writes:

>Everyone on this list is quite different.

It would be interesting to see what basic interests and views the members
of
this list hold.  For a few people, published works answer this pretty
clearly but that's not true for most list members.

I'll start.

I'm a dilettante.  I am most interested in spending my time wondering how
small a chunk of "code" could implement AGI.  I believe that the physical
and cultural structure of a human being's environment provide the vast
majority of the complexity of a human mind, and there may be a rather
small
amount of complexity (necessary "code") needed to leverage that
environment.
  I'm interested in figuring out exactly what this small core would need
to
do.

I guess something like 10^16 operations working on 10^16 bits of
information
is roughly what's required for "human equivalence", given operations and
bits defined roughly in current computer terms.  I think those numbers are
just about average for those interested in AGI who have made public
guesses.
  Since there are still so many orders of magnitude between any hardware I
can get my hands on and the amount required, I'm content to just think
about
things for now and don't feel a big need to write lots of code right now.

I do think it is crucial for us to build AGI before it can run at human
equivalence in real-time, so we can understand it and hopefully avoid
scenarios like Yudkowski describes where self-improvement rapidly spirals
out of control.  If it turns out that a Commodore 64, Pentium4-based pc,
or
BlueGene are already powerful enough for AGI, we could be in trouble.

What about the rest of you, what are your interests?


-----
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


-----
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/?member_id=231415&user_secret=fabd7936

Reply via email to