Re: AI progress ( or lack thereof)

From: Dani Eder (
Date: Thu Dec 09 2004 - 16:07:24 MST

> AI research itself has had a poor
> track record, so why do you
> assume future AI research will be more prodigious
> than future nanotech,
> cryonics, biomedical research, etc.?

Hans Moravec has estimated the processor power
required for AI is 10 Tflops. Until a few years
ago no machine was that powerful, and right now
no machine used for AI research (as far as I know)
is that powerful yet. So AI progess has been
held up by lack of suitable hardware.

I believe that if you plot the complexity of the
AI software measured in lines of code or similar
metric vs processor power required to achieve AI,
there will be some curve below which human-equivalent
AI can't be done, and above which it can.
Unfortunately, we have very little idea where that
curve lies.

At one extreme, you could have a program that is
as simple as:

For NEURON = 1 to 10^11
  Update state of NEURON

Where NEURON is a sufficiently detailed model of
human neurons to behave similarly.

I call this the 'brute force' approach, and my
wild-assed gues is that it would require on the
order of 3000 Tflops to run at human real-time
rate, but I wouldn't be surprised to be off by
a factor of 30.

At the other extreme, the size of the software would
dominate the processor power costs when trying to
configure a system (RAM and Disk storage) or would
create data transfer bottlenecks. So beyond some
point it dosn't make sense to try to save processor
power with better software.

An example at the other extreme would be a lookup
table that answers every question that can be asked.
Huge storage requirement, but not very much processor
time to answer any given question.

Looking at developing an AI from another angle, the
human Genome and a Windows XP installation CD have
about the same data storage capacity (~700 MB). The
human genome generates a human intellect, so in theory
a piece of software of similar complexity should
be able to generate an AI. but

But no AI project has come within several orders of
magnitude of the resources Microsoft has used to
develop their operating system, and their product
is still a long way from having the 20 year uptime
it takes a human to mature.

So besides inadequate hardware to date, another
explanation of AI research's poor track record is
inadequate software development resources.


Do you Yahoo!?
Jazz up your holiday email with celebrity designs. Learn more.

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:50 MDT