Re: Failure of AI so far

From: Samantha Atkins (samantha@objectent.com)
Date: Sun May 25 2003 - 23:08:37 MDT


On Saturday 24 May 2003 03:21 pm, Ben Goertzel wrote:
> Perhaps this is just semantic nitpickery, but I don't think it's
> accurate to call it a *programming* bottleneck.
>
> It's an AI design bottleneck.

I believe that both are a problem. Certainly the field of software
engineering generally bears out the fact of programming bottlenecks.
Within the limits of finite human brain power and our tools and
procedures to date there are real limits to what can successfully be
tackled. I don't think it unimportant to acknowledge those limits.

>
> And, until this AI design bottleneck is convincingly solved, we
> won't know for sure if AI requires more processing power than is
> currently available. We also won't know if there is a programming
> bottleneck -- i.e. depending on what the correct AI design is, AI
> may or may not require vast numbers of lines of code.
>

LOC is of course a crude measure of software complexity. For
instance, parallelism is one of the types of software problem that
humans generally are not that good at understanding, implementing and
maintaining. AGI, even in seed form, will be the most massively
parallel system ever contemplated. I would expect at the least that
some new techiniques and tools would need to be developed to cope.
Some of these could turn out to be byproducts with financial
possibilities.
 
> I'm betting that our Novamente design breaks through the AI design
> bottleneck, but I don't pretend to have adequate evidence of this
> to convince skeptics.
>
> And, the Novamente design is pretty well fleshed out, but even if
> you believe it, we don't understand the dynamics of a completed
> Novamente (which we don't have) well enough to tell if it will need
> 200 or 20,000 state-of-the-art Linux boxes to achieve a kick-ass
> level of general intelligence...
>

Very interesting. The lower figure can easily be hit by a small
company. The upper figure is 3 google linux box farms worth.
Pricey but doable. If these figures are reasonably accurate then
almost the entire problem is design and implementation bound.

- samantha



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:42 MDT