From: Matt Mahoney (matmahoney@yahoo.com)
Date: Fri Nov 28 2008 - 06:43:06 MST
In my AGI proposal at http://www.mattmahoney.net/agi2.html I estimated the cost of AGI to be on the same order as the value of the human labor it replaces, US$1 quadrillion. This is a very optimistic estimate. It assumes a million-fold increase in computing power at negligible cost, an infrastructure of pervasive public surveillance, and solutions to the language and vision modeling problems. Under these assumptions, the dominant cost is programming the AGI with 10^17 to 10^18 bits of knowledge, which I estimate is the complexity of the global economy, i.e. how much knowledge is needed to do what people currently do to earn money. This estimate is based on 10^10 people each with 10^9 bits of knowledge, as estimated by Landauer, and 90% to 99% knowledge overlap between people. To estimate overlap, I consider the cost of replacing an employee (sometimes a year's salary) as the amount of knowledge that can't be copied from somewhere else.
The vast majority of knowledge needed to program AGI is in human brains. The internet currently has about 10^11 web pages with perhaps 10^4 bits each, only .1% to 1% of what is needed. The rest has to be extracted from human brains at a rate of 2 bits per second per person. The cost of AGI is dominated by this requirement.
For this reason, I specified natural language for communication between peers and surveillance for gathering data. As for the language and vision modeling (needed to implement peers), most of this (at least for language modeling for data compression) is currently done in C, C++, and assembler for efficiency reasons. Almost none is done in Java. Language and vision are both highly parallelizable, so I expect that languages that support parallelism will dominate. This might be done on top of existing languages, for example, CUDA or Google's MapReduce on top of C++.
-- Matt Mahoney, matmahoney@yahoo.com
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:01:03 MDT