RE: Hardware Progress: $212/GFlop/s

From: Ben Goertzel (
Date: Tue May 06 2003 - 17:20:14 MDT


In my view, your perspective is a bit naive in terms of cognitive science.

To explain why fully would take more than a single e-mail. But the main
point is that most of the knowledge that a mind requires is not of the sort
that can easily be articulated in a file to be loaded into an AI system.
Most human knowledge is:

a) highly detailed tacit/implicit declarative knowledge about familiar
things, usually unconscious in humans


b) procedural knowledge, detailed knowledge about how to do particular
things in particular contexts, nearly always unconscious in humans [how to
prove a theorem... how to talk to a girl.. how to walk down the street ..
how to phrase a sentence based on the education level of the listener ..
etc. etc.]

This kind of knowledge has got to be learned -- at least, until we figure
out how to read detailed knowledge out of human brains. Or until we create
the FIRST real AI system, then we can copy it with variations ;)

The well-known Cyc project is an attempt to encode all human knowledge in
explicit form, to be loaded into an AI brain. Many AI experts, myself
included, think this is doomed to fail, for the reasons stated above (though
it may fail interstingly, and may lead to valuable results short of Real
AI). However, we do have on this list Stephen Reed, one of the leaders of
the Cyc project, who is of a different opinion!

-- Ben Goertzel (,

> -----Original Message-----
> From: []On Behalf Of Tim
> Duyzer
> Sent: Tuesday, May 06, 2003 6:49 PM
> To:
> Subject: Re: Hardware Progress: $212/GFlop/s
> I don't see machines needing 2.5 years to train. Training time
> for humans is
> a mix of social training and education. Unlike machines, though, we can't
> simply load data into a brain. Give a superintelligent machine access to a
> resource such as the Internet (which by then would likely be far more
> extensive and powerful than it is today, perhaps completely
> unrecognizeable
> at all) and it would take very little time for it to become adept
> at social
> interaction as well as extremely knowledgeable - it would have the ability
> to assimilate and process information on many more levels than a human. In
> the end, there'd really be no need for any kind of AI training - just plug
> it in.
> Tim
> > The processing power to run a human-level AI is
> > unknown, but is estimated to be in the range of
> > 100 - 100,000 TFlop/s. Assume the most powerful
> > machine available for AI work is in the $3 M price
> > range. Then the expected date of a human-level
> > machine is 2008-2025. The expected date for
> > superintelligent machines is 7.5 years later. This
> > is calculated from a human taking about 20 years
> > to train. A machine 8 times faster than a human
> > is expected 5 years after a human-equivalent, and
> > it would be expected to train in 20/8 = 2.5 years.
> >
> > Daniel

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:42 MDT