From: Damien Broderick (d.broderick@english.unimelb.edu.au)
Date: Mon Apr 15 2002 - 08:33:25 MDT
Here's a draft chapter from my current book, a collection of essays on the
need for a recovery of Enlightenment values, possibly to be titled KNOW-ALL
(in the USA that might well be KNOW-IT-ALL). The idea below is meant
seriously. I've run it past an active specialist in quantum computation,
who finds it sound but feels that-- But we'll deal with that in the
afterword, below.
Please do not repost this anywhere else without my permission.
===================
Quantum AI Bootstrap, a 1/2-Baked Idea
I had a rather appalling idea recently, based on my short story `Infinite
Monkey'. If it works, the idea could change the world as we know it. I know
at least one man who is trying to make it work.
But first some background: according to a widely accepted interpretation
of quantum theory, every choice and event that ever happens is literally
doubled and reduplicated, with small crucial variations, in trillions upon
trillions of diverging parallel realities, spread through infinite lateral
time.
If you are having a bad hair day, cheer up--another version of you, in a
universe at right angles to this one, is doing just fine. And a zillion
more, it's true, take up every conceivable alternative position in between.
Some of them are dead. Some are on Mars. A few are sharing Graceland with
Elvis, who is married to Princess Di or perhaps John Lennon.
The many worlds hypothesis was proposed in 1957 by Hugh Everett, III, who
in this universe died decades ago. According to one of his followers,
Oxford physicist David Deutsch, this outrageous version of reality ought
not surprise us, or at least physicists, since it is simply the best theory
available to science--quantum mechanics--taken perfectly literally without
metaphysical evasions. The multitude of overlapping histories is needed to
explain some of the most basic features of our ordinary world. And there
are some striking and unexpected consequences of this idea. Deutsch argues
ingeniously, for example, that gene sequences performing the same function
in different adjacent worlds must be closely similar, while so-called `junk
DNA' littering genomes will vary at random. It is precisely this
consistency in true informational structures that marks them off from noisy
rubbish, however complex and elaborate the noise might seem.
Deutsch is one of the founders of quantum computing, and no New Ager. His
extravagant ideas are worth taking very seriously indeed. My own notion
might deserve no such respect, but I like it anyway. It generated itself
spontaneously over good Japanese food and ample wine in the ingenious and
quick-witted company of James Newton-Thomas, a mining robotics engineer,
and our companions. James explained a nifty way to use quantum theory to
search for mineral deposits. I thought it was very clever, and I hope it
works and makes him (and ultimately the rest of the human species)
enormously wealthy, so obviously I am not going to tell you about that.
Recalling my own whimsical fiction, I thought perhaps we might be able to
create a self-bootstrapping AI--artificial intelligence--that phones home
once it congeals out of the quantum noise of the infinite overlapping
universes that exist, according to Many Worlds quantum theory, jammed
together all around us, unseen but not without their impact upon us.
Here is what you do to create this wonderful machine.
Set up a printer driven by, or coupled to, a true quantum randomizer (a
radioactive isotope emitting purely random electrical noise, for example).
This becomes, in effect, a quantum die--and such dice have their equivalent
in every universe that is henceforth entangled with our own.
Start by generating a random alphanumeric list of a million characters,
each chosen by a throw of the quantum die. A very simple program does that
easily.
If it were written in an appropriate code, the listed output from this
process would be a computer program able to run on the computer sitting in
front of you. Here, now, is the key step: at the end of the space reserved
for the alphanumeric gibberish, add a clear instruction for the program to
locate you in superspace and get in touch with you as soon as possible.
In most of the stacked-up, entangled universes, this randomly generated
`program' will be nothing but junk mail. In some, it will contain the
entire text of Hamlet, in others it will read out as a macabre cookbook for
making scones from worms; in a few, the words it contains will be so
piercingly poignant that the version of you reading them will weep and vow
to change your life for the better.
But in at least one superposition, we might hope--if such an outcome is
mathematically and computationally not impossible--the jumble of utterly
arbitrary letters and numerals will comprise, by pure chance, a valid AI
program suitable for the computer at hand (either a standard desktop, or a
military intelligence hypercomputer, depending on who is initiating this
stunt).
In short, your computer tries to compile and run code in every superposed
universe. In most, the pseudo-code--the noise, the copy of Hamlet--just
sits there on the hard drive or disk doing nothing. In a few, it might be
destroyed by a giant meteorite impact, or eaten by the dog.
And in at least one superposed cosmos, the AI program runs, wakes up--and
dutifully calls home.
Calls every home, presumably.
Would it be Friendly? Almost certainly not. (Not that it would necessarily
be *unfriendly*, mind you.)
Will someone do this experiment anyway? If it's doable, surely.
Will the AI sit there--before or instead of calling `home'--and bootstrap
itself up to godhood? Well, hell, I would.
Could happen any day now.
======================
The objections and my responses:
1) >You could try running your machine over and over again until the cows
come home, or the big crunch occurs. Or am I missing your point somehow?
I think so. It depends on how smart the self-bootstrapping code becomes.
The key to success is the bit at the end of the program, which is an
injunction (or politely phrased quid-pro-quo request) for the new AI to
find a way to phone home with the news and then to do so.
2) >What do you mean "find a way to phone home"? You mean to communicate
with the "other branches" of the wave-function?
Just so.
>Unfortunately, this does not appear to be possible, in general.
Ah, but this is just *our* current understanding of QT, which is
acknowledged to be incomplete. The possibility entwined in my little notion
is that the randomly-generated AI will quickly Spike itself into
hyperintelligence and find a way to do it by hook or by crook. On the other
hand, the fact that it hasn't spontaneously happened already makes it look
as if it's never going to--unless in those worlds, and only those, that
remain tightly entangled with the initiating project, wherever someone
tries it.
====================
Damien Broderick
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:38 MDT