From: Bill Hibbard (email@example.com)
Date: Wed Dec 11 2002 - 02:17:29 MST
> > Basically, this is an intro into why I think it is very difficult (nigh
> > impossible) to build an AI that is superior than a human (across
> > the board).
> > There isn't a clear target to aim at, or a meaningful measure of success.
> This strikes me as a rather silly argument. The fact that "superiority" is
> subjective is no reason that we can't build AI's that have vastly superior
> intelligence to humans in many meaningful senses.
> Are cars and trains "better" than horse-drawn carriages? Are planes
> "better" than birds?
> Is human intelligence superior to dog intelligence? Perhaps dog
> intelligence is superior in some senses. It may be superior at finding
> other dogs to mate with, for example....
> The vagueness of the word "superior" should be dealt with by proposing
> particular, contextually-meaningful measures of superiority.
> I can define a lot of meaningful measures of "superior intelligence."
> For example, if I can create a program that can:
> 1) Prove any mathematical theorem from the math research literature (without
> being shown the proof)
> 2) Figure out how to modify humans (via genetics and/or pharmacology and/or
> nanotech) so that humans don't age.
> 3) Create a molecular assembler a la Dreyfus
> 4) Create original music that the majority of humans find extremely
> emotionally moving
> 5) Effectively manage a team of people working on a project via e-mail and
> chat communication
> 6) Interact with humans in a way that makes many humans feel positively
> spiritually touched
> 7) Formulate a unified theory of the four known physical forces
> 8) Defeat any human in all games of mental skill
> 9) Perform surgical operations on humans with more skill than any human
> 10) Operate mobile robots that drive cars, fly airplanes and spaceships, and
> climb mountains
> then I will rate this program as having "superior general intelligence" to a
> human. Perhaps you don't agree with this list of goals [of course, the
> details could be modified], but it's certainly clear and meaningful....
In addition to the skills you listed, I think good, practical
measures for machine intelligence will be:
1. How many humans can a machine know well?
2. How many humans can a machine converse with simultaneously?
Humans can know about 200 other humans well, so there is plenty
of room to improve on human intelligence. Of course, the skills
you listed will make intelligent machines interesting and valuable
companions. My machine companion will be more valuable to me if
it personally knows almost all other humans (e.g., "There's a guy
in Burma you should meet."), so there will be a motive to maximize
this measure of machine intelligence.
This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:41 MDT