Re: Genius Dogs

Hal Finney (
Thu, 9 Oct 1997 11:02:36 -0700

Thanks to John Clark for summarizing Drexler's plan for evolving super

> A recipe for intelligence: Build a simulated world in your
> computer and fill it with very simple creatures (programs).
> Make sure they must solve problems in order to get "food". The
> creatures that are better at solving problems leave more
> descendants. Now you do nothing, just step back and let it
> evolve. After evolving for a few hundred million SIMULATED years
> you have intelligence, high order intelligence.
> How long would it take in real years? He calculated the amount
> of computer power needed to simulate ALL the brains that have
> ever existed before humanity, that is, all the brains since
> brains were invented in the Cambrian Explosion 570 million years
> ago. He concluded that 10^38 machine instructions would do the
> trick. A Nanotechnology computer the size of a large present
> day factory and using no more power, could perform 10^38
> machine instructions in about 2 years.

Without seeing Drexler's original proposal, it is hard to evaluate this in
detail. But taking it literally as presented, there is an obvious flaw.
Just simulating the *brains* which have existed would not be enough to
evolve intelligence. Brains don't just become intelligent on their own.
What makes them become intelligent is the challenge presented to the
brains by the environment. Brains which evolve in such a way that they
are able to better survive and reproduce will become more numerous.

So it would be necessary to simulate not only the brains, but the bodies
and the environment where the brains interact. We'd have to simulate
competition, the challenges of survival. Frogs need tongues and flies
to evolve brains which will try to predict the flies' motion. Apes need
hands, trees, branches, vines, and a reason to jump from tree to tree to
evolve brains which can let them leap through the air and catch a branch.

A brute force technique, then, would have to simulate not only the brains,
but potentially a significant subset of the biosphere. This would be many
orders of magnitude more difficult. Volume considerations alone would
suggest 4 or 5 orders of magnitude (what fraction of the biosphere is
made of brains?). A complicating factor is that the kinds of simulation
necessary or appropriate for things like muscles, trees and rivers may be
fundamentally different than the presumably neural-net based simulations
we would prefer to use for brains (and which probably formed the basis
of Drexler's estimate).

The idea of evolving intelligence does seem to have potential, and may
turn out to be the easiest way of developing AI. I'm sure that Drexler's
example is not meant literally, but is intended to motivate the intuition
that given nanocomputers, it should become practical to evolve AI. This
may still be the case even if the example doesn't go all the way towards
establishing the point.