Re: Media update (evolving AI in Cyberspace)

Brian Atkins (
Tue, 23 Dec 1997 12:11:28 -0600

Hal Finney wrote:
> Brian Atkins, <>, writes:
> > On Nightline tonight, they were discussing future technology,
> > including uploading via neuron replacement and evolving AI.
> > They showed some cool video of what appeared to be evolving
> > GAs in a "real" 3D environment. This was the first time I'd
> > seen that, and unfortunately I didn't catch the person's name
> > who had done it. Basically they started with a simple organism
> > composed of two cubes, and let it evolve into successively
> > more complex organisms based on how well they could "swim"
> > through the simulated environment. The result looked very
> > similiar to a sperm cell :-)
> Very interesting! One problem with such simulations though is that
> sometimes they are "cooked" so that a particular outcome is more or less
> inevitable. There may be implicit constraints on the types of forms the
> genetic algorithm can evolve, and the authors will have been careful to
> set things up so that a sperm-like shape would be one of the possible
> outcomes.
> Could they take that same simulation, make the fluid thinner (like
> air) and change the fitness function to evolve something that flies?
> That would be a test of whether they had a good general model of evolution
> or just a toy.

Yes, later on after that they showed how they changed the
constraints to a "land" environment. They dropped the best
swimmer (the sperm) on it and let it start wiggling. After
many generations they had several good "walkers" evolved,
but they mentioned that many of them were quite different
than what they had expected to end up with.

> > I couldn't help but think about the
> > possibilities should a 3D Cyberspace ever replace the web-
> > would it become possible to evolve AIs by letting them grow
> > and learn within such a system?
> Theoretically, other "geometries" for cyberspace would not preclude the
> evolution of AIs.
> I've been reading Dan Simmons' science fiction series Hyperion recently.
> (I can't really recommend it overall because much of it is terribly
> violent and sad.) He has an AI society called the TechnoCore which at
> first I assumed was an outgrowth of 21st century AIs. However later
> in the series he claims that it actually evolved out of alife parasites
> like our computer viruses and worms.
> The problem with this idea is that it is hard to see how or why alife
> would evolve intelligence. Current viruses are about as smart as natural
> viruses, probably. And it's not clear how much it would help them to be
> more intelligent. If you added today's AI technology into a virus, like
> a fuzzy logic based expert system to choose a strategy for infiltrating
> a new computer, it would probably not be successful. The virus would
> become so bulky that it would be hard to remain undetected.
> Maybe in the future things will change, computers becoming so big and
> sloppy that smarter alife can find a niche. Then it may be possible to
> see a real form of alife evolution where it smartens at least in the
> biological sense, where a bacterium can be considered quite "intelligent"
> in its adaption to its environment.
> Hal

Well, I don't see how a virus could ever do anything unless they
build some sort of genetic algorithm into it that lets it mutate
when it reproduces, or even trade bits with other viruses.

But if you had a nice Cyberspace that actually created a sort
of real world then things would be interesting. You would have
an environment with light, sound, gravity perhaps, and most
of the other things that we have here on Earth, plus live
people (avatars) interacting inside it and tons and tons of
data (the whole web would presumably be incorporated inside,
plus original content developed for Cyberspace). Now imagine
someone releases some software that allows people to use the
spare cycles on their computers to run an organism inside
Cyberspace, and the organism is allowed to mate with other
people's organisms using various genetic algorithms (perhaps
mutate the algorithms too?), and then the best of the offspring
(best for Cyberspace) is kept. So you could have thousands of
these things running around, and with sufficient algorithms
and time then they should become more and more complex.

I haven't really thought much about this, but it just seems
that a good Cyberspace would be a much better environment for
this than trying to do it on the web/inet.