Re: rambling question on AI

From: Eliezer S. Yudkowsky (
Date: Fri Jan 04 2002 - 16:50:14 MST

Reason wrote:
> From my readings online, there's
> very little that discusses what happens in between
> a) useful primer steps like Cyg [Cyc?], Flare, etc
> b) boom! functioning, if very crude general intelligence or seed AI
> Which I would assume is because no-one really knows for sure (or else we'd
> be hip deep in AIs by now).

I would say that a "functioning, but very crude GI" *is* an intermediate
step, if by "functioning but very crude" you mean "the modules are talking
to each other but the overall system doesn't do anything interesting".

To get an "intermediate application" before this point requires that you
try and implement something less than a GI. And to do this, you are going
to have to do different work in different ways than you would if you were
taking a direct path to general intelligence. You are going to have to
expend a lot of effort on building a functioning system only some of which
will be useful as a way station to GI. I would be afraid of getting
bogged down in something that wasn't directly on the road to GI - but
then, SIAI is a nonprofit and thus has different strategic requirements;
we can try for seed AI and accept intermediate applications whenever they
show up, rather than having to convince VCs about the shortsightedness of
destructive shortcuts. Peter Voss or Ben Goertzel may have different
goals and consequently different plans.

I have to say, though, that one of the largest reasons - maybe even *the*
largest reasons - why SIAI is a nonprofit is that general intelligence and
seed AI is pretty much one of the most complicated things in the world,
and if there's a VC with voting control sitting there and saying "Make
this simpler so it can go out the door in six months", then AI is just not
going to happen there.

> Now if you're walking into a VC's office to pitch an AI company -- one of
> the few places you're guaranteed not to be the strangest project they've
> seen -- they're going to want to hear how you're going to be generating
> revenue within six months from an initial investment. [Since you're a
> software company, that is, and producing those easy-to-make software things,
> as opposed to an infrastructure company making those hard-to-make lumps of
> iron, but I digress -- lambasting VCs is far too easy].

Six months? A dream.

> So you're starting out with something along the lines of a concept, perhaps
> some useful protocode for classifying information and information
> processing, possibly even a grasp of the One True Path To AI. What can you
> produce in six months that you can sell? What is the improvement on that
> that would be coming out in a year?

In six months, you can produce a toy program that is totally off the
development path to general AI. I think that's pretty much it.

-- -- -- -- --
Eliezer S. Yudkowsky
Research Fellow, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Fri Nov 01 2002 - 13:37:32 MST