Thank you, Matt.
Well then, "{L]ike everyone else on this list, I do not have a working AGI"
(Mahoney, 2018). However, I've also been thinking, reading, researching AI for,
in my case, about 30 years. In 2011, I put some of my thoughts on paper and
submitted them in the form of 2 papers to the AGI conference: one briefly
explaining my proposed high-level architectural principles and the other one
had a (my) proposed solution for bootstrapping it i.e grounding it in reality
without relying on robotics or having to hand-code huge knowledge bases or
teaching all of our knowledge to it. Sadly both papers were rejected.
Given 'the call' and also because I hold some of the earlier posters (Matt,
Alan, Rob, Steve, Jim, etc) in especially high esteem, I now attach the
architectural one. {Footnote: I still feel a bit sore about the rejection;
apart from having wanted to get audience feedback and attend the conference, my
normal acceptance rate for conference papers is well above 80% and I've
published well over 200 peer-reviewed academic papers so far}.
Since the rejection, I've not really worked on a prototype. But as soon as I've
got another $300-$500K together, I can take early retirement and resume my
thinking, tinkering 😉 and hopefully work properly on a (Python) PoC.
Obviously my thinking has evolved further since then and I've also obviously
got a lot more details and notes (I actually had a detailed list of minimum
required modules and their functions), but those are not in a publishable
format. Critical feedback is always welcome.
Jean-Paul
________________________________
From: Matt Mahoney via AGI <[email protected]>
Sent: Saturday, 09 June 2018 6:55 PM
To: [email protected]
Subject: Re: [agi] Anyone interested in sharing your projects / data models
Like everyone else on this list, I do not have a working AGI. It is
easy to underestimate the enormity of the problem. The most obvious
application of AGI is to automate human labor. Globally this is a USD
$75 trillion per year problem. A working solution would have a ROI of
world GDP divided by market interest rates, about $1 quadrillion. When
investors won't even put in $1 million into your project, they are
effectively setting odds of a billion to one against your success.
When you count the number of false promises and failures to solve AI
since the 1950's, that's not unreasonable.
It's not that I haven't tried. Ten years ago I proposed an AGI
composed of billions of narrow AI specialists and a distributed index
to route your requests to the right agents. Building this would
require a global effort over decades and an economic infrastructure
that rewards providing useful services in a hostile and competitive
distributed computing environment. Agents would compete for attention
and reputation in a world where information has negative value. The
distributed index provides a message posting service, where all
messages are public, signed and dated and cannot be edited or deleted
once posted, and are routed to anyone who might care. You can find the
proposal at
http://mattmahoney.net/agi2.html<https://protect-za.mimecast.com/s/EG-sCj2gl8i3P4PVhnJyr3>
Such a project is far beyond what one person or even a large company
could accomplishment. A human brain sized neural network with 10^14
synapses and 10 ms clock would require 10 petaflops and about 1
petabyte of RAM. Such computers exist but require over 1 megawatt of
electricity, compared to about 20 watts for the human brain. You would
need 7 billion of these computers to replace 7 billion workers,
requiring 7000 terawatts of power. Global energy production is
currently 15 TW. This kind of power reduction is not going to be
achieved by further shrinking transistors, which are already close to
the limits of physics. It will require computing by moving atoms and
molecules rather than electrons, something that biology has already
figured out but is still a long way off for us.
Perhaps there are more efficient solutions to AGI. Sure, for some
problems, like arithmetic. But deep neural networks are still the best
we have for vision and probably language. The problem is that
intelligence (measured by reward or prediction accuracy) over a wide
range of problems increases logarithmically with CPU time and memory
because the problems themselves have a power law distribution over
difficulty. This clearly shows as the economy grows linearly while
computing power grows exponentially. It shows in my own work in data
compression as a measure of natural language prediction accuracy. The
two graphs in
http://mattmahoney.net/dc/text.html<https://protect-za.mimecast.com/s/BHhfCk5jm7Sr1y1PHk_vfN>
are 10 years old but
the relationship hasn't changed.
Ambitious, decades long AGI projects like OpenCog and NARS are
knowledge representation frameworks with empty knowledge bases and
inadequate hardware to solve any useful problems even if they were
populated. Why? Because software is easy and human knowledge
collection is hard. The design complexity of the human body cannot
exceed the compressed information content of our DNA, which I
estimated to be 300 million lines of code. (These numbers from my 2013
paper "The Cost of AI" published in 2017 as a chapter in "Philosophy
of Mind: Contemporary Perspective" (Curado, Gouveia eds).
http://mattmahoney.net/costofai.pdf<https://protect-za.mimecast.com/s/sSbeClOkn7IApwpjFguJLN>
). This is doable for $30 billion,
a tiny, insignificant fraction of the total cost. Knowledge collection
is much harder. Human long term memory capacity is 10^9 bits, of which
1% is unique to you and can only be extracted by speech or typing at 7
bits per second costing $5 per hour at global average wage rates. This
is only practical (and still costing hundreds of trillions to collect
10^17 bits) by giving up privacy and publishing your personal data.
Otherwise you are stuck with endless surveys and repeating the same
data over and over. How many times today did you have to type in just
basic info like your name or email?
Hoping to stimulate some intelligent discussion.
--
-- Matt Mahoney, [email protected]
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T731509cdd81e3f5f-Mb45fd1f76d6ed4014a05cc4f<https://protect-za.mimecast.com/s/vwItCmw0oyUADoD9F4Bukx>
Delivery options:
https://agi.topicbox.com/groups<https://protect-za.mimecast.com/s/9RfFCnZmpOtKjZjkcPjoBe>
Disclaimer - University of Cape Town This email is subject to UCT policies and
email disclaimer published on our website at
http://www.uct.ac.za/main/email-disclaimer or obtainable from +27 21 650 9111.
If this email is not related to the business of UCT, it is sent by the sender
in an individual capacity. Please report security incidents or abuse via
https://csirt.uct.ac.za/page/report-an-incident.php.
------------------------------------------
Artificial General Intelligence List: AGI
Permalink:
https://agi.topicbox.com/groups/agi/T731509cdd81e3f5f-M2677cb2f097c49d2353bfdf2
Delivery options: https://agi.topicbox.com/groups
AGI11 Yapaam Yet Another Proposed Architecture for Artificial Mind.docx
Description: AGI11 Yapaam Yet Another Proposed Architecture for Artificial Mind.docx
