Re: Moravec's estimates?

From: Brian Atkins (
Date: Sat Apr 07 2001 - 20:11:48 MDT

Jimmy Wales wrote:
> James Higgins wrote:
> > I, too, find it difficult to believe that a game console will leapfrog
> > super-computers and be the first to reach 1TF.

Well the Playstation 2 did run into supercomputer export law restrictions,
and is rumored to be used in Iraqi missle guidance systems...

> I don't think that's what is being claimed. We are already at
> 12 teraflops with the fastest supercomputers, and there is every
> indication that we'll be at 1000 teraflops by 2005.
> In Moravec's terminology (MIPS), that is that we are at
> 12 million MIPS now, and will be at 1 billion MIPS less than
> 5 years. (A billion million is 1000 trillion.)
> Now, if Moravec is right and the raw power of the human brain is
> 100 million MIPS, then we're getting really close, at least as far
> as really expensive supercomputers go.
> On the other hand, if Kurzweil is right and the raw power of the human
> brain is 20 billion MIPs, then we've got a lot further to go -- 10-15
> years extra, we might suppose.

Actually if those teraflop CPUs from Sony/IBM pan out in 2006, it should
be possible to build a "Blue Gene 2" that would have around 30 petaflops.
(a Blue Gene CPU is only 32gflops)

Or even if not, if Moore's law is doubling at at least 12 months by the
time Blue Gene is done, then it would only take 4 or 5 years (2011 ish?) to
get to the 20 petaflops range. I think 20 petaflops by 2010 is pretty
realistic if the supercomputer curve keeps on the trajectory it has
been on lately.

> I wanted to see if anyone knows of any good reasons to say that
> one or the other of them is closer to right.

Most cognitive scientists I hear from seem to think it may not be so much.
These are the scientists who believe we can understand enough of the human
brain to be able to simplify it into less computationally intensive
algorithms. However, I think what is likely is that it will take an
overshoot of CPU power... because the initial real AI software will not
be anywhere near perfectly optimized. So the first real success will take
many times more power than is really absolutely necessary. Which is good
I guess since it will give the AI plenty of room to improve itself even
on its starting hardware.

> OR, maybe Penrose is right, and there are quantum aspects to human
> cognition, so we need quantum computers to even come close to the
> performance. But I don't know of any solid biological or philosophical
> reason to think Penrose is right.

Well lately the progress in quantum computing seems even faster than
Moore's law, so no worries :-)

> But, I'm no expert.

Me either :-) Who IS an expert on this? At this stage everyone has their
own opinion. It won't be until we get near having a real working AI that
we will be able to know for sure what it needs. I will be interested to
see how intelligence scales with how much MIPS/memory is available. In
the future will we measure intelligence not by IQ but by the size of our
computing substrate? :-)

Brian Atkins
Director, Singularity Institute for Artificial Intelligence

This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:36 MDT