James Rogers wrote:
> The plasticity of the hardware is utterly irrelevant. It is the interaction
Agree here. Not quite, actually, because nanotechnology could introduce
a substrate that actually restructures itself adaptively, dynamically.
For instance, one could implement some rarely changing control circuits
in bucky circuitry, which would be more efficient than computronium
for that specific, rarely but changing task.
> and complexity of the *data* that matters, which is why one should be able
I'm kinda missing the word "speed" here.
> to do AI on just about any reasonable piece of silicon. There is no magic
Hmm. Reasonable as to "number of bits" and "speed of tweakage of said bits",
yes. Silicon does imply photolithography, and photolithography does
imply a certain structure size, flatland, and Moore saturating on that
substrate in about a decade.
So, I disagree with the silicon part of "reasonable piece of silicon".
I would rather use the more neutral word "substrate".
> in evolvable hardware, it is simply a fast way of evolving data interactions
> that could be done completely in software on a boring von Neumann machine.
Yes, but an AI which takes a MYear to say "Hello, world!" is perhaps not
very recognizable as an AI, especially given that you can't assess the
AIness of a precursor, if it's impractically slooooowwwwwwwwwwwwwwwwwwwwwwwwwww.
If I want to replay one GYear of evolution in practical time, I obviously need
something which is a great deal faster than realtime.
> Hardware and software are equivalent things; hardware is faster, software is
> cheaper, and you balance the two depending on the specifications of the
> project at hand.
Basically, you don't balance very much, for most current purposes hardware
is a constant. If you want the most of crunch these days it means investing
in air conditioning and a hall full of PCs, and a large number of fat switches.
> Give me just one example of something you can do in high-plasticity
> evolvable hardware that can't be done in software.
Speed, of course. That was easy.
This archive was generated by hypermail 2b30 : Mon May 28 2001 - 10:00:03 MDT