Re: [sl4] what's with all the math?

From: Pavitra (
Date: Wed Oct 14 2009 - 16:34:53 MDT

Luke wrote:
> - I may not have specified levels correctly on that task in version 0.
> When I said "When will we know we have succeeded?" I should have said "When
> will we know we have succeeded in creating a friendly AI?", and not "When
> will we know we have succeeded in creating design requirements?"

That's what I thought you meant.

> - I'm also a web developer, and I'm pretty confident that at least some
> aspect of our future strong AI will be written in code. However, I predict
> that the code will be simply managing tasks which are handled by black-box
> systems like neural networks.

The neural networks will be written in code.

> - Mathematical definition of friendliness - good luck with that. Perhaps
> we could start with English definitions, then slowly start constraining our
> language more and more until we can get a formal mathematical definition.
> Does anyone know of a body of formal mathematical definitions of
> sociological interactions?

Economics and game theory might be a good place to start.

> I think there's a big tradeoff between rigor and applicability. Linear
> systems (cannon firing at � from horizontal with velocity V - where will the
> cannonball land?) lend themselves nicely to mathematical analysis - i.e. we
> can answer the question without performing the experiment. Nonlinear
> systems (6 * 10 ^ 23 concepts interconnecting and mutually
> activating/inhibiting within a Jupiter Brain - will it try to kill us?) not
> so much - i.e. we won't really know until we build the darn thing and turn
> it on.

The cannonball is only a simple system because we have a good high-level
model of the 10^whatever quantum interactions that make it up. My
understanding of the FAI question is that it's trying to produce a
similarly high-level model of AI systems that will allow us to predict
the overall behavior of a Jupiter brain without explicitly tracking
every flip of a bit.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:05 MDT