Re: Building a friendly AI from a "just do what I tell you" AI

From: Hector Zenil (hzenilc@gmail.com)
Date: Sun Nov 18 2007 - 18:01:04 MST


There is another potential issue concerning questions about models of
AI, both very likely to happen as I will explain below:

1. The problem of OAI deciding if a model (created by itself or not)
is FAI is undecidable.
or
2. it is irreducible.

in both cases OAI cannot give a definite answer. Now replace OAI by
anything else but a hypercomputer.

(2) implies (1) but not the other way around. i.e. (1) is the hardest
subset of the type (2).

By Rice theorem we know that (1) is very likely --with measure 1-- and
(2) follows also immediately and independently of (1) since FAI is
supposed to be as sophisticated as OAI or any AI (including HI, the H
for human, under computationalism).

Rice theorem says that any nontrivial property about the language
recognized by a Turing machine is undecidable. A property of Turing
machines is trivial if it holds for all partial computable functions
or for none. So asking whether or not some AI model is an FAI model is
evidently non-trivial because there is at least one Turing machine
that has the property, and at least one that hasn't.

(2) says that if the system is as sophisticated (namely universal) as
the "tester" then there is no possible shortcut to decide a property
of the other model (running also over a universal Turing machine). (2)
can be read as a problem of computational complexity, but it does not
always follow from that field since computational complexity concerns
about time and space but not steps. This latter is considered by
another type of interesting irreducibility to Stephen Wolfram. His
principle of computational irreducibility basically says that while
some computations admit shortcuts that allow them to be performed more
rapidly, most do not. And by his principle of computational
equivalence, it would turn out that any descent approach to AI,
including FAI, will turn to be computationally irreducible. In other
terms, this latter principle says that there are no natural (feasible)
intermediate Turing degrees so if a computational system is sound and
computationally powerful enough it will follow immediately that it is
universal and therefore irreducible.

Hector Zenil

On Nov 19, 2007 12:01 AM, Thomas McCabe <pphysics141@gmail.com> wrote:
>
> On Nov 17, 2007 5:46 PM, <sl4.20.pris@spamgourmet.com> wrote:
> > Building a friendly AI(FAI) from a "just do what I tell you" AI(OAI ==
> > obedient AI).
> >
> > I know that OAI have been discussed recently in this forum but read on
> > before you dismiss this.
> > To avoid any possibility of dangers we program the OAI to not perform
> > any actions other than answering with text and diagrams(other media
> > like sound and video would be a possibility too). In essence what we
> > would have is a glorified calculator. I think this avoids any dangers
> > from the AI following orders literally with unintended consequences.
> >
> > So we go to the OAI and say: "Tell me how I can build a friendly AI in
> > a manner that I can prove and understand that it will be friendly."
> >
> > The OAI will think and give you a detailed blueprint, proof, etc...
> >
> > You then analyse the documents until you understand them. You could
> > also ask for further clarification from the OAI.
> > Someone might raise the objection: how can you be sure that there
> > aren't any backdoors or problems with the blueprints? This will also
> > be a problem if you come up with your own way of making a FAI. The
> > only answer is: you have to be very careful! The point of using an OAI
> > is the same as for using a calculator: to make things easier.
> >
> > Then you build the FAI.
> >
> > Of course the real thing may be a bit more complicated for example:
> > making the OAI first generate plans for a more intelligent OAI and so
> > on. We could have several OAI enhancement steps until we finally are
> > able to make a FAI.
> >
> > On a very basic level our current-date computers are OAIs.
> >
> > Comments?
> >
> > Roland.
> >
>
> The OAI analyzes your instruction, concludes that it would require a
> lot of computing power to design an FAI, and then turns the planet
> Earth into computronium before spitting out the design plans.
>
> - Tom
>

-- 
Hector Zenil-Chavez
hector.zenil-chavez@malix.univ-paris1.fr
Université de Lille I (Laboratoire d'Informatique Fondamentale)
Université Pantheon-Sorbonne -Paris 1- (IHPST)
--------------------------------
zenil.mathrix.org
animaexmachina.com
---------------------------------
Fondation Suisse
Cité Internationale Universitaire de Paris
7, bd Jourdan - chr. 114
75014 Paris
France
--------------------------------------------------------------


This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:00 MDT