[ExI] Hard Takeoff-money

[Samantha Atkins]

On Nov 16, 2010, at 8:39 AM, Keith Henson wrote:

> On Tue, Nov 16, 2010 at 5:00 AM,  Samantha Atkins <sjatkins at mac.com> wrote:
> 
>> On Nov 15, 2010, at 7:31 AM, Keith Henson wrote:
> 
> snip
> 
>>> What does an AI mainly need?  Processing power and storage.  If there
>>> are vast amounts of both that can be exploited, then all you need is a
>>> storage estimate for the AI and the average bandwidth between storage
>>> locations to determine the replication rate.
>> 
>> But wait.  The first AGIs will likely be ridiculously expensive.
> 
> Why?  The programming might be until someone has a conceptual
> breakthrough.  But the most powerful super computers in the world are
> _less_ powerful than large numbers of distributed PCs.  see
> http://en.wikipedia.org/wiki/FLOPS

Because:
a) it is not known or much expected AGI will run on conventional computers;
b) a back of envelop calculation of equivalent processing power to the human brain puts that much capacity, at great cost, a decade out and two decades or more out before it is easily affordable at human competitive rates;
c) we have not much idea of the software needed even given the computational capacity.

This leads to quite high  likelihood that the first AGIs will be very expensive.  

> 
>> So what if they can copy themselves?  If you can only afford one and they are originally only as competent as a human expert then you will go with entire campuses of human experts until the costs comes down sufficiently - say in a decade or two after the first AGI.
> 
> The cost per GFLOP fell by 1000 to 10,000 in the last decade.

That is relevant but not determinative of early AGI cost.

> 
>> Until then it will not matter much that they are in principle copyable.    Of course if someone cracks the algorithms to have human level AGI on much more modest hardware then we get lots of AGI proliferation much more quickly.
> 
> Any computer can run the programs of any other computer--given enough
> memory and time.  The human brain equivalent can certainly be run on
> distributed processing units since that's the obvious way it works
> now.

You are assuming that an AGI runs on a general purpose computer.  This may be false.  It would require a massive fine grained parallel processing for instance or such great speed and throughput as to fully simulate such.    Any turing machine may be able to run any program but that doesn't mean that it can run it well enough or fast enough to have any real benefit whatsoever.

- samantha