RE: Alien science

From: Ben Goertzel <>
Date: Sat, 30 Nov 2002 21:44:41 -0500

Tim May wrote:
> It will be interesting and exciting if you are right, but I think the
> kind of AI you describe above and below is further off than 10-30
> years, though perhaps not 50 years.

Well, clearly neither of us has a rigorous way of making an exact

But my main point stands regardless of whether it's 10, 50 or even 100

> I have no idea. True, it may come up with all sorts of weird theories.
> But, absent new experimental evidence, will these new theories actually
> tell us anything new?

I suspect there is a lot of "new evidence" to be found thru reinterpretation
of the raw data observed using current experimental apparatus, including
observations not categorized as "events" but rather as "noise."

And I suspect that an AI with a different conceptual framework would be able
to milk a lot more information out of variants of the basic apparatus we use
now, without moving to thousand-mile-long accelerators.

> (Needless to
> say, given that today's best AI programs and computers are having a
> very hard time even doing "naive physics," a la ThingLab and its
> descendants, I'm not expecting progress very quickly. And ThingLab is
> more than 20 years old now, so expecting massive breakthroughs in the
> next 10-20 years seems overly optimistic.)

Given the exponential rate of advance we're seeing in many kinds of
technology, there is certainly a clear potential for progress in this area
to be tremendous during the next 10-20 years. But, of course, history shows
that there's no easy way to predict in advance which aspects of technology
are going to exponentially advance at a given time, and which will stagnate.
AI has stagnated lately whereas, e.g., computer hardware and brain scanning
have exponentially increased.

> More importantly, would an AI version of classical physics, complete
> with incomprehensible (to us) phase spaces and n-categories and so on,
> including constructs with no known analogs in our current universe of
> discourse, would this version give any predictions which differ from
> our own? In short, would the AI's version of physics give us any new
> physics?

Well, that depends perhaps on what you mean by "new physics," I think.

Right now our physics is basically stumped by most complex systems. We
resort to

-- computer simulations
-- crude "phenomenological" models

I suspect that an AI could come up with a real "science of complex systems",
identifying "probably approximately correct" patterns describing the
behavior of complex systems.

These complex-systems patterns would be consistent with, but qualitatively
different from, the predictions made by human physics.

> Probably so. But at any given point in time the best we can do is to do
> the best that we can. We of course cannot just wait for the machines...

No, but we can try to create them -- my current endeavor ;)

> Whether there "are" branes and strings and spins and suchlike at the
> Planck scale is unknown to me, but physicists seem to be making
> progress in acting as if such things have some meaning.

The amount of progress being made by theories about branes and strings is
highly debatable, actually.

Not all physicists consider this stuff "real physics", and I feel there is
some justification to their skepticism.

>But the fact that reality is so weird is not, in my opinion,
> an argument of any kind that we should not try to make some sense of it
> with the best arsenal of tools and concepts we can gather.

Certainly, I think trying to make sense of the microworld is a noble &
intriguing pursuit, or else I wouldn't be on this list!!

> > Will mathematics be central to this new physics? Maybe. But not our
> > mathematics.
> I disagree fairly strongly on this point. I think our mathematics is
> what is most lasting, albeit the mathematical ideas change names and
> new ideas become more important.
> And I expect the mathematics the AIs develop, or that alien cultures
> may already have, will look like a "coordinate transformation" on our
> space of mathematical basis vectors. Their categories may be slightly
> different, but the underlying structure will be similar.

I really doubt it.

I spent a long time trying to make existing math concepts say something
sensible about complex systems.

I failed.

This is not just my personal failure either -- many other smart folks have
tried and failed as well...

I think a fundamentally new mathematical direction will be needed to make a
mathematics that can truly handle complexity, in the micro or macroworld.

String theory, for instance, merely creates unmanageable complexity, it
doesn't provide a conceptual framework for dealing with it...

> > Anyway, this was part of why I decided to start thinking about AI
> > rather
> > than fundamental physics ;->
> AI remains interesting, but I think new views of physics will be coming
> from AIs long after other important things come out of AI. Just my
> opinion.
> (For example, some friends of mine are doing interesting work on using
> systems of several million machine agents to data mine large amounts of
> financial data. It seems likely that this kind of work on machine
> learning, pattern extraction, support vector machines, and a plethora
> of other "AI tools" will have major effects on the world of economics
> and forecasting. And on creating financial derivatives (synthetics)
> which are alien to human thinkers/investors.)

Yeah, financial forecasting with AI does not require Artificial General
Intelligence (AGI) in any sense, it is a classic domain-specific narrow-AI

Whereas, coming up with new physics will require a significant degree of
general intelligence, I believe.

In this sense, physics theorizing is certainly a much harder problem than
financial prediction-- it's hard to argue with that!!

I tend not to even consider that kind of narrow-AI work "AI" -- I just think
of it as computer science. But I have to remind myself periodically that
the mainstream of academia does consider this AI, and considers AGI work to
be a foolish and faraway dream...

-- ben g
Received on Sat Nov 30 2002 - 21:44:15 PST

This archive was generated by hypermail 2.3.0 : Fri Feb 16 2018 - 13:20:07 PST