Re: Lilliputian Posthumans

Eugene Leitl (
Fri, 23 Oct 1998 12:38:44 +0200 (MET DST)

John Clark writes:
> Difficult yes, impossible no. Hans Moravec figures it would take 10^13
> calculations per second to emulate a human brain. To build one of Drexler's
> nano computers that could perform 10 ^13 calculations per second you'd
> need about half an ounce of matter, mostly of carbon, and about 15 watts
> of energy.

A modern high-end PC delivers roughly 10^9 ops peak. Such performance is distinctly underpowered to run even _one_ neuroanatomically correct neuron in realtime. Since we have several 10^9 neurons to account for, I do not see how a factor of 10^4 can make a measurable difference.

Even if we drop the requirements of having to emulate a particular human brain and settle for an intelligent system instead, I very much doubt that a 10 k node 21264 Beowulf is capable of intelligent action, however carefully programmed. (Thinking of several billion transistors/node I can't help but to notice that a 10 k node Beowulf would seem a particularly wasteful way of utilizing them, maybe a 10^14 transistor WSI FPGA++ run by GA could make a fair score on the Turing test).

I don't think that Moravec's notorious underestimates are doing the whole field of AI a credit.