Dear Eugen,
Has any thought been made among the AI community as to how to prevent
the constructed AI from being Solipsists? How does one AI deal with the
existence of other AI? Don't they have the same Turing Test problem, but
even more accutely given the thesis of hardware independence?
Kindest regards,
Stephen
----- Original Message -----
From: "Eugen Leitl" <eugen.domain.name.hidden>
To: "Lee Corbin" <lcorbin.domain.name.hidden>; <everything-list.domain.name.hidden.com>
Sent: Wednesday, June 29, 2005 5:26 AM
Subject: Re: People, Machines, and Manipulations
On Tue, Jun 28, 2005 at 03:51:46PM -0700, Lee Corbin wrote:
> > An instantiated program is much more than a sequence of
> > bytes -- it also has state. Most programs do not have much
> > state, but some (AI, specifically) are completely dominated
> > by state.
>
> Well, so are people. I am actually in a state of bewilderment at
People have lots of state, and interestingly the current electrochemical
activity (the spatiotemporal pattern of spiking) is only a minor part of it.
The pattern of a person is readily regenerated from a flat-EEG lacune, so
the
state is encoded as neuranatomy, not as electrochemical activity of the
neuroanatomy. At the very least, we still consider those the same person.
> the present moment at how you are using some words :-)
It's not entirely my fault, the common use of words like "program" to
describe activity of a physical object (a person) is somewhat misleading.
> Very unusual choice for meanings of words. On your usage, the
> chess program... oops, there isn't a single chess program! I
A chess program pressed on a CD is a single program, regardless how many
copies are pressed (provided, no errors occur). Multiple instances of a
chess
program running perfectly synchronized are still one program, compared bit
by
bit at the same clock tick (very fast very large systems run into hairy
relativistic issues, in regards to clocks and comparisons). Once they're
allowed to deviate, they become distinct individuals.
Now this is quite silly in case of a chess program, because it doesn't
represent a very complex world. It is rather close to a genetically
homogenous population of C. elegans, which are all very close to being the
same
archetypal worm, since their neuronal connectivity is genetically
determined.
If you have a behaving complex animal it has to represent a lot of
information about itself, and its environment.
> mean to say that on your usage of terms, after the chess
> software (if I may) plays 1. e4 it becomes a different program
> after I reply 2. Nf3. I better avoid using the word "program"
> if we are to communicate!
We have to settle for somewhat cumbersome but more precise terms like
individuals (a static frame snapshot of the state), and a similiarity metric
over the space of individuals, and evolution trajectories over the state
space of individuals (several subsequent static frame snapshots). If you
look
at a molecular dynamics program dumping trajectory frames it's exactly the
same.
> Hopefully, I can refer to what I want as a Turing Machine, and
> you won't pull the rug out from under me by saying that each
> time it goes into a new state, it's a different Turing Machine.
Yes, and no. The device is the same, but it encodes different individuals
depending on state of its tape. The computer is the same, but it can run
quite a large number of different programs.
> > Biology doesn't make a clean distinction between software and hardware.
> > I agree there is similarity/homology between me-former and me-today,
> > but that similarity is difficult to measure at a low level.
> > Synchronizing
> > spatially separate discrete systems and make measurements on bit vectors
> > is
> > something relatively simple, at least in gedanken.
>
> Yes. So what do you think about the possibility of uploading?
> That is, transferring your entire intelligence and values from
> its present biological substrate to a silicon-based one. Do you
Silicon doesn't compute very well, so I would prefer a more generic
"in machina" to an "in silico" (which is about to become as archaic as in
"in relais" or "in tubus vacuo").
> consider it possible that technology from the year 3000 (were it
> somehow applied to where you are at this moment) could transform
No problem, assuming cryonics works.
> you into a robot who didn't know that the transformation had
> taken place, and yet you would then consist of a system where
> there was a clean distinction between hardware and software?
Absolutely (though there are considerable difficulties present in mapping a
scanned slab of neuroanatomy voxels to a slab of computing molecules; this
involves automatic feature extraction and hierarchical model building by
machine learning, and is quite beyond the state of the art in modelling).
The hardware is simple enough (a molecular electronics/spintronics CA
would suffice, and probably be even optimal, though you'd probably need a
mole of switches for a low-level simulation of a human primate), but
there is going to be an awful lot of state.
For all practical purposes you can consider a person a very, very large bit
vector, making only sense in the right computational context, of course.
--
Eugen* Leitl <a href="http://leitl.org">leitl</a>
______________________________________________________________
ICBM: 48.07100, 11.36820 http://www.leitl.org
8B29F6BE: 099D 78BA 2FD3 B014 B08A 7779 75B0 2443 8B29 F6BE
Received on Wed Jun 29 2005 - 14:54:57 PDT