Re: Loosemore's Proposal

From: Michael Wilson (
Date: Mon Oct 24 2005 - 18:44:50 MDT

> Richard Loosemore wrote:
> There is a REASON why one should build a reasonably complete system,
> watching the learning mechanisms *really* build concepts (and not rely on
> you the programmer to insert them) and watching the recognition system
> *really* figure out the relationship between sensory patterns and
> recognized internal concepts (rather than have you assume that this can be
> done later.....

Well of course. I don't see how you could reasonably propose to build
an AGI without doing this. The critical difference between an opaque
design and a transparent one is that in the later you can understand
exactly what is going on instead of guessing. The critical difference
between constructive design and trial-and-error is that in the former
you implement to test specific hypotheses and go back to the drawing
board to refine your model if it turns out that they're wrong, while
in the latter you simply search for things that appear to work,
without having a preexisting (or in many cases, post-discovery) model
that explains why they work.

> have substantial empirical data that says that if you invent your
> local mechanisms and don't test them in the above way, those
> mechanisms may simply not EVER work, no matter how much fiddling you
> do with them.

Obviously theory doesn't actually /do/ anything until it's
implemented, and I agree that it's generally impossible (for humans)
to be sure whether a design will exhibit the desired performance
without building and testing it. Note that questions of capability
are /not/ primary goal of FAI theory, which is concerned with
ensuring that the effects of an AGI on its environment will be of a
certain type, and makes no guarentees about whether a design will
actually be effective at achieving goals given some finite amount of
computing power. That question must be answered (IMO) by rather
different means, where implementation and experimentation have an
important role, but in a transparent and constructive context rather
than an opaque and/or trial-and-error one.

 * Michael Wilson

How much free photo storage do you get? Store your holiday
snaps for FREE with Yahoo! Photos

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:52 MDT