Ed Porter
From: Edward Porter [mailto:ewporter@gmail.com]
Sent: Thursday, 14 July 2011 8:20 AM
To: AGI
Subject: [agi] WHAT DON’T WE KNOW HOW TO DO IN AGI?
Before I go to AGI '11 in little less than a month I would like to do a survey on this list to find out what, if
any, conceptual problems stand between us, and being able to achieve human level AGI.
In this survey, I am assuming that well before the end of this decade thousands of AGI researchers will
each have access to computers with at least 1 million cores, one peta ops per second, 100 TB of
memory, 10 trillion random read-modify-writes per second into that hundred terabytes, and a theoretical
global bandwidth of 20 trillion messages per second (which I am assuming translates into the ability to
effectively use approximately one trillion global messages a second). For want of a better name, I will
refer to such hardware as a "mega-tera machine". It is my belief that if industry focuses on making such
machines starting within two or three years, there is a two out of three chance that by the end of the
decade they can be manufactured for well under $100,000 apiece, and profitably sold for under
$200,000. (And systems of 1/10th to 1/100th this power could be sold for under $20,000 and $2000,
respectively, enabling tens of thousands of AI researchers to experiment with them.)
So I am asking, what hard conceptual problems stand between us and achieving human-level AGI if we
assume we have such hardware. I am asking you to set aside the "small hardware" mentality that has
dominated AI for most of its existence, and to instead think big and imagine how you might design a
human-level AGI's assuming that you have something within at least several orders of magnitude of the
power of the human brain, itself.
I am NOT interested in:
(1) problems that relate to trying to get AGI's to operate on hardware substantially less powerful than the
mega-tera machine I have described above
--- but I would be interested in reasonable arguments as to why roughly human level AGI could not be
created on a platform having the hypothesized capabilities of the mega-tera machine.
(2) problems relating to the large number of design choices, parameter tunings, and trial and error
explorations of design spaces that will be required to piece together efficiently and automatically
functioning human level AGI's from the tremendous wealth of knowledge and concepts we already have
about how to build AGI
--- but I would be interested in reasonable estimates of the the difficulties for the AI community to solve
most of them within, say, five years after, say, 1000 groups have access to mega-tera machines for
purposes of AGI research --- given the expected advances in AGI that will be made between now and
near the end of this decade when such machinery could become common.
I will be interested in hearing your comments, and from time to time I will try to prepare summaries of the
reasonable problems that have been so suggested.
Thank you.
Ed Porter
From: Edward Porter [mailto:ewporter@gmail.com]
Sent: Thursday, 14 July 2011 8:20 AM
To: AGI
Subject: [agi] WHAT DON’T WE KNOW HOW TO DO IN AGI?
Before I go to AGI '11 in little less than a month I would like to do a survey on this list to find out what, if
any, conceptual problems stand between us, and being able to achieve human level AGI.
In this survey, I am assuming that well before the end of this decade thousands of AGI researchers will
each have access to computers with at least 1 million cores, one peta ops per second, 100 TB of
memory, 10 trillion random read-modify-writes per second into that hundred terabytes, and a theoretical
global bandwidth of 20 trillion messages per second (which I am assuming translates into the ability to
effectively use approximately one trillion global messages a second). For want of a better name, I will
refer to such hardware as a "mega-tera machine". It is my belief that if industry focuses on making such
machines starting within two or three years, there is a two out of three chance that by the end of the
decade they can be manufactured for well under $100,000 apiece, and profitably sold for under
$200,000. (And systems of 1/10th to 1/100th this power could be sold for under $20,000 and $2000,
respectively, enabling tens of thousands of AI researchers to experiment with them.)
So I am asking, what hard conceptual problems stand between us and achieving human-level AGI if we
assume we have such hardware. I am asking you to set aside the "small hardware" mentality that has
dominated AI for most of its existence, and to instead think big and imagine how you might design a
human-level AGI's assuming that you have something within at least several orders of magnitude of the
power of the human brain, itself.
I am NOT interested in:
(1) problems that relate to trying to get AGI's to operate on hardware substantially less powerful than the
mega-tera machine I have described above
--- but I would be interested in reasonable arguments as to why roughly human level AGI could not be
created on a platform having the hypothesized capabilities of the mega-tera machine.
(2) problems relating to the large number of design choices, parameter tunings, and trial and error
explorations of design spaces that will be required to piece together efficiently and automatically
functioning human level AGI's from the tremendous wealth of knowledge and concepts we already have
about how to build AGI
--- but I would be interested in reasonable estimates of the the difficulties for the AI community to solve
most of them within, say, five years after, say, 1000 groups have access to mega-tera machines for
purposes of AGI research --- given the expected advances in AGI that will be made between now and
near the end of this decade when such machinery could become common.
I will be interested in hearing your comments, and from time to time I will try to prepare summaries of the
reasonable problems that have been so suggested.
Thank you.
Ed Porter
 | |  | |  | |  | |  |