Re: "Real" programming and seed AI

From: Dani Eder (
Date: Thu Feb 07 2002 - 21:21:14 MST

> why shouldn't we get
> > the latest PC or supercomputer, start hacking in
> self-modifying
> > assembly for the first generation of seed AI?

Design and optimization of complex systems is an
engineering discipline where I work (Boeing), and
it happens to be the one I have spent my career
doing. The answer to the above question depends
on what your goals are. If the goal is to instigate
a Singularity as soon as possible, the right course
of action may not be to start coding right away.
It might be to work on AI architecture and
algorithms, because clever solutions of such will
allow you to use smaller hardware, and thus
afford the required hardware sooner.

Goals can be organized under the headings of
Cost, Schedule, Performance, and Risk. They
can also be sorted into factors and drivers.
factors have an additive effect on the 'goodness'
of your technical approach, while drivers have
a multiplicative effect. For example, a
performance goal for a seed AI might be 'able
to improve itself by learning or self-modification
at a rate of at least 100% per year'. The
rationale for the 100% figure is that the
hardware is improving at that rate, so a slower
growing seed won't be able to utilize the
improving hardware effectively. So you might
assign a scoring function where 50% per year
gets a score of 0, 100% per year gets a score of
5 and 200% per year gets a score of 10. You
then feed the goal scores for all your technical
approaches into a scoring function of the

D(1) x D(2) ... D(n) x (F(1) + F(2) + ... F(n)).

where D(i) is a driver score and F(i) is a factor
score. The approach with the highest score is the
one to choose.

The point of this method is to quantify the goals
to help you choose the best way of going about meeting
the goals.


My point is if you haven't specified what the
goals are, how can you evaluate whether your
technical approach is a good one or not?

Do You Yahoo!?
Send FREE Valentine eCards with Yahoo! Greetings!

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:37 MDT