From: Robin Hanson

Message-ID: <>
Date: Tue, 5 May 1992 16:51-0700
Return-path: <>

Lillipution Uploads

(sci.nanotech submission)

Many people are repelled by the idea of uploading, in part because the usual descriptions of it are very alien -- we become abstract computer-based software entities living in arbitrary virtual realities. An uploaded human brain running on a desktop computer might run many millions of times faster than us ordinary folks. At such speeds the time penalty for slowing down enough to experience an ordinary body would be far too great to allow many such indulgences. People's strong need for the familiar physical sensations and comforts would have to satisfied in a virtual reality that had little direct connection to ordinary physical reality.

But the future will still have a lot of physical work to be done, work which could greatly benefit the incredible space/time reasoning and control abilities our brains have developed over the ages. Perhaps a million slower minds would often be more economically useful than one super-fast one. I imagine people will want to make their new uploaded world as similar as possible to the old familiar world, just as we now try to keep close to nature and try to preserve familiar ways of dressing and organizing households. Though they could not overcome overwhelming economic pressure, such social tendencies can be remarkably strong.

Thus I am drawn to imagining lilliputian uploads -- tiny bodies shaped like ours with a brain cavity filled with not much more computer power than it takes to keep up the faster body movements of their smaller bodies. This approach allows us to speed up thinking, devote a substantial fraction of our mass to advanced brain hardware, and yet continue our direct interaction with physical reality.

While various body sizes would be possible, I suspect standardization pressures would encourage one (or at most a few) standardized body sizes. Thus I see small cities landscaped with real but small plants, houses and churches build from wood, all looking recognizably like the world we live in. As now, wealthier citizens would be able to better afford country estates that look more like the old days, and the less wealthy may settle for more functional and temporarily-alien accommodations.

Of course many things would be strange compared to now, with backups, travel by "teleportation", and the wrenching economic, social, and legal changes resulting from the ability to make copies. But this would be all the more reason for people to cling to the familiar.

In fact, I have done a nano-calculation and estimate these uploads would stand within an order of magnitude of a 1/4 inch tall (the same size as in "Honey, I Shrunk the Kids" :-), and run perhaps a few hundred times faster than us (with much more uncertainty in this figure). Thus billions of these uploads might live in one current office building.

Here is my calculation:

Assume we are just scaling down a 6 foot tall person using similar materials, so that the periods of typical body motions scale with the body size. Our 1500 cc brain cavity is scaled down also, and filled with 1/2 fast nano-mechanical RAM, and 1/2 with nano-mechanical CPUs. The specs of these devices are conservatively estimated by Drexler [upcoming NanoSystems book] to be: CPU: 10^9 instructions per second (= 1000 MIPS) fits in a cube 400 nm (nanometer) on a side, uses 90 nW (nanoWatt). Contains 10^6 "transistors" and 10^4 bit registers.

RAM: as fast as CPU registers takes 40 nm^3 per bit stored.

Our current brains use about 25 Watts to have our estimated 10^15 synapses fire about 10 times per second each. Let us say that simulating a brain takes c*10^15 bits of memory, and s*10^15 instructions per second, where we may differ on estimates of the variables c and s. The smaller c and s get, the better our software will have become at compacting memory and speeding up the simulation.

Merkle ["Energy Limits to the Computational Power of the Human Brain", Foresight Update #6] estimates c < 1, s = .001 to 10. Moravec [Mind Children] estimates c = .1, s = .01. Schwartz [special AI issue of Daedalus] estimates c = 100, s = 1000. Results below will be parameterized in terms of c and s.

Let me also introduce a parameter x describing how fast we run the nanomechanical CPUs relative to their nominal speed. Slower speeds generate less heat. O.K., here are my estimates:

height:                  l =         c^1/3 * 6.8 mm
body motion speedup:     b =        c^-1/3 * 266
mind speedup wrt body:   m =   c^4/3 * x/s * 2.4  
cpu heat generation:     h =       c * x^2 *  56 Watts
total speedup wrt us:    t =       c * x/s * 640

For uploads that live in air and don't have to drag a cable around with them, 56 Watts seems a bit much - they would glow like Tinkerbell! If the mind is to at least be able to keep up with the body, we need some wins in s relative to c the allow us to lower x. That or accept an upload brain dominated by CPU. Note that x << 1 would require using reversible software (which avoids erasing registers), whose extra overhead would cost another constant penalty of ~ 10.

Note that if c = .4, one of our days would be a year for them, like living in the arctic circle is for us.