[extropy-chat] Extreme Intelligence

[Lee Corbin]

Russell writes

> > Could an intelligence exist using our basic human architecture
> > that could rapidly solve problems far harder than anyone can
> > solve today?

> What do you mean by "basic human architecture"? Hardware or software?

I explained later in my post: limited to the same senses. Also,
though, I should have indeed said something like: consisting
of a single cerebral cortex. I definitely want to exclude some
kind of arbitrary multi-minded computronium-based solution,
because probably no one could argue that *that* would not be
possible.

> > Here is what I have in mind: suppose first that there is a
> > canonical way to extend the IQ scale. 

> But the IQ scale itself isn't canonical. Even among humans,
> who all have basically the same hardware and software architecture,
> there isn't a single measure - IQ is, for example, a notoriously
> bad predictor of things like wisdom and social skills.

Of course. But as a measure of g, that is, of cognitive ability,
it serves astoundingly well.

  * figure out how a fusion energy reactor could be designed, 
    and write up specifications sufficiently detailed so that
    the rest of us could build the thing

  * provide a specific outline of how an AGI could be coded-
    up following the outline/design by a good software team 
    in six months

  * be able to understand at about normal reading speed any
    book ever written, much as you can easily understand and
    absorb everything being conveyed by a Dick-and-Jane book

> No. The above tasks would require _knowledge_ - whether a
> particular CAD design would make a viable fusion reactor,
> for example, is a fact about the real world, not about symbols
> shuffled inside the mind.

Well, yes. Thanks, this *is* the kind of criticism I solicit.
But it'll take a little more to persuade me; after all I am
positing *extreme* intelligence. No upper bound, (except as
some possible limitation to our architecture).

You are, of course, quite right about knowledge. But the rate
at which knowledge can be acquired is surely commensurate with
intelligence, wouldn't you say?

As for extending IQ, recall how chess and some other straight-
forward tasks work. In chess, you stand the same probability
of defeating a player 100 points below you regardless of your
talent or level of skill (this is how the Elo scale was
devised, and how scales in East Asian sports, for example,
the famous Dan and Kyu, work).

By considering a huge variety of tasks, perhaps it becomes
possible to describe a level of intelligence X such that X
is to 200 as 200 is to 180, and then to call that 220. In
other words, we take enough samples of tasks demanding
cognitive capability, and calibrate our measure accordingly.

(By the way, your examples of wisdom and social skills simply
don't correlate with cognitive ability as measured by 
practitioners in the field; they're pretty much in agreement
on this.)

> And to ascertain that fact will require more than a week of
> work, irrespective of how fast the mind proposing the design
> thinks.

I wish I knew how you are so confident. Are you saying that
it's flat out impossible no matter how smart a piece of matter
is?  Or one limited to, as I say, "human architecture"?  I must
hasten to say that by allowing a week (or a month, whatever),
I am supposing that lengthy and costly experiments do not need
to me done, and I could be entirely wrong. Somehow, it seems
strikingly interesting and peculiar if it emerged that no matter
how smart some piece of matter was, all it could do would be to
shrug and say "the data simply is not available, and you would
have to do the following 26 experiments...".

> The same is true of intelligence; it simply isn't a mathematical
> function of an information processing system. There just isn't
> any quantity of which an increase will enable a system to deliver
> answers to arbitrary problems. (You can use one of the measures
> based on Kolmogorov complexity that have been proposed, but if
> you do you'll find it doesn't correlate well with performance at
> real world tasks.)

Why isn't it a mathematical function of an information processing
system?

Lee