RE: Real vs. Simulated Ecologies [was: Moravec article available]

From: Billy Brown (
Date: Wed May 03 2000 - 13:25:48 MDT

Robert J. Bradbury wrote:
> The work by Bennett, Landauer, etc. say the "cost" in computing is
> the cost of erasing information. For example, if I have a capacitor
> storing some electrons, there is a heat produced in removing those
> electrons to discharge the capacitor. They way you reduce this problem
> is to allow the capacitor to charge and discharge very slowly (along the
> ideas of reversible computing). However in a virtual world it isn't
> clear to me that those constraints exist. If you have the freedom
> to design the ecology, could there be a way that allows the
> virtual erasure of some information to result in the virtual
> creation of other information (along the lines of allowing
> the electrons to cycle between two capacitors). You have
> to design the system so that the transforms on the information
> waste less energy than similar transforms using "real" atoms
> or molecules. For example -- can I design a virtual system
> where the energy cost of breaking a virtual C=C bond is cheaper
> than the energy cost of breaking a real C=C bond (~1.19 aJ)?

I think we have some level-crossing confusion going on here.

The thing you have to remember about a VR is that it isn't an alternate
universe. Any transformation that happens in the virtual world has to be
implemented as an actual transformation on real particles in a physical
system. It doesn't have to be the *same* transformation, but it has to be
*some* transformation of equal or greater information content. No matter how
cleverly you design the VR, it will always have this constraint due to the
simple fact that the virtual world has no independent existence.

Now, of course you can simulate an event for less energy than it takes for
the real event to happen. We do that all the time, and the relative
efficiency will only grow as we find better ways to do the simulations. But
the reason we can do that is because we discard all of the information we
aren't interested in, instead of keeping it all. What you want is a way to
have extra bits in the VR that don't exist anywhere in the physical world,
or "free" computations that don't have to be performed by the underlying
system. Unfortunately, neither of these things is possible.

> Ah, but there are many examples, say from mathematics, where a simple
> formula can represent a huge amount of information. Moravec's article
> points out how optimizing compilers can transform one representation
> of a process into a much more efficient representation. The Transmeta
> efforts show how you can design a machine that can refine processes
> into increasingly efficient operations.

Yes, but any transformation that could be used to improve the efficiency of
a virtual computer can also be used to improve the efficiency of a real one.
Or, to put it another way, a physical computer can use arbitrarily complex
and abstract operations. Making a simulated computer doesn't get you any
benefits you couldn't have gotten with a suitable abstraction layer, and it
imposes a lot more overhead.

> I agree there always has to be a physical representation. But almost
> all "information" in the current "real" world is represented in atoms,
> atomic bonds, electric charges in specific locations, etc. I'm thinking
> along the lines of a "virtual" computer where the information is
> represented by photons at specific positions in space. It seems
> likely that a computer of that type (i.e. no matter, other than that
> required to produce the photons), could potentially have a much higher
> information density than "matter" based computers or Moravec's
> Higgsinium or Monopolium (that are still based on "atomic" models).

If photonic computers are the way to go, the efficient approach is to
actually build one. Running it on some other kind of computer as a
simulation doesn't get you anything - simply running a computation is always
more efficient that running a simulation of a computer doing the same

Billy Brown

This archive was generated by hypermail 2b29 : Thu Jul 27 2000 - 14:10:26 MDT