From: Brian Atkins (brian@posthuman.com)
Date: Tue Jun 18 2002 - 21:41:44 MDT
Anders Sandberg wrote:
>
> On Tue, Jun 18, 2002 at 12:06:27AM -0400, Brian Atkins wrote:
> >
> > You're missing my point, which is that something as "super-smart" and "super-
> > fast" as you were talking about should be able to make the VR realistic
> > enough (hence my use of the word "emulate") to completely do any needed
> > testing in that arena. Or do you believe uploading is impossible? The
> > same level of emulation capabilities are required in either case- if it
> > can do one it can do the other it would seem?
>
> I run across people from time to time who think that since I spend my
> time studying simulated brains instead for real brains, we could do away
> with all animal experiments and use simulations instead. It doesn't
> work, since to make a good simulation you need the data of the
You need the data once (at most, assuming you can't deduce it in software
from prior results), and after that you can simulate it to your heart's
content. This is why we see more and more moves to simulate more and more
items of engineering interest in software. I just read a couple days ago
about a proposal to begin trying to do a complete human using grid
computing. Once you get enough data to fully understand something there
really should be no more reasons to keep fiddling with reality in the
vast majority of cases.
> experiments. Even if we had simulations so good that they seemed just
> like the real thing, how can we be sure they really depict the real
> thing and not just something close but subtly different? I certainly
Because eventually your simulation tools get developed to the point of
being perfect. This may or may not end up applying to certain things in
quantum mechanics, depends on what an SI can do with a quantum computer,
or if it ever figures out the simple Wolfram CA underlying the Universe :-)
> them and that always leaves room for surprises; the emulation of the
> entire Earth would require some kind of scanning that interacted with
> every atom in the planet - which in itself is close to the kind of
> potentially devastating tech we are discussing the deployment of in the
> first place; computronium isn't arbitrarily cheap, and to *emulate* the
> Earth you would need a computronium system of comparable mass. So it
> doesn't seem that an emulation is feasible.
Well remember we are talking about the technology needed for a fast
Singularity, not the tech required to build a Dyson sphere. There are
relatively few bits of tech required for a seed AI to be able to get
really smart, and make everyone else really smart. I don't see why
accomplishing those two tasks would require emulating the whole planet...
it might only require emulating something the size of a human, or a
roomful of equipment.
>
> Simulations are just arguments in the form of "given these assumptions
> and initial conditions, X will happen". You can run validation by
> changing assumptions and initial conditions, which is good, but you
> won't get certainty. A good simulation will give you a small uncertainty
> and help convince you and others that a risk may be worth taking, but it
> is still not an empirical argument. Physics is constantly throwing upp
> weird effects that doesn't show up in simulations, even in apparently
> well understood areas **. If you start testing genesis machines based
> solely on simulations, please warn me beforehand so that I can flee the
> area. It is just the recipe of a Hollywood disaster.
I don't think this argument applies to a SI which will likely be able
to come up with a final TOE, etc. etc. But anyway going back to the
original point, this kind of potential problem does not seem to prevent
a fast Singularity. Building some nanotech safely using the Foresight
guidelines, and building some computronium, and developing an uploading
process for volunteer humans are not things in the same league as a
Genesis device, and therefore do not require the kind of utter absolute
certaintly that it will work the first time without errors.
-- Brian Atkins Singularity Institute for Artificial Intelligence http://www.singinst.org/
This archive was generated by hypermail 2.1.5 : Sat Nov 02 2002 - 09:14:54 MST