Foresight Update 12
page 2
A publication of the Foresight Institute
 Xerox
PARC Update
Modeling an assembler on a computer before trying to build one
seems a sensible idea. It's faster and cheaper, and it lets us
look beyond current manufacturing limitations. Of course, an
assembler might have millions of atoms. If we want to model such
a structure on a computer, we'd have to enter the coordinates of
those atoms. Many molecular modeling software packages use a
"point and click" interface to add atoms to a
structure. If we want to specify the location of each atom, we'd
have to press the mouse button millions of times...
This horrible image inspired the obvious thought: automate the
problem. Write programs that generate the coordinates of all the
atoms in the structure. If the program accepts a few parameters,
then any one of an entire class of structures can be generated
quickly and easily.
We've already written the first such program: a little more than
800 lines of C generates any member of a class of tubular
structures. Conceptually, the tube is made simply by taking a
flat rectangular sheet of diamond and bending it until the edges
meet and can be bonded together. (This is a description, not a
recipe for making the structure.) The length, radius, tube wall
thickness, and the crystal surface and orientation can all be
specified by the user. Crystal surfaces are usually specified by
a triplet of integers, such as: 111, 110, 010, 322, etc. By
entering a triplet, the nature of the surface of the tube wall
can be specified. Finally, the rotational orientation of the
surface with respect to the axis of the tube must be specified.
Appropriate changes in this parameter can make the
"grain" of the surface point along the axis, at right
angles to the axis, or spiral down the length of the tube.
With a fast and easy method of generating tubular structures,
it's easy to investigate many of the questions that naturally
arise. For example, as the tube wall becomes thicker and thicker
and the radius becomes smaller and smaller, the strain becomes
greater and greater. Finally, when the strain is too great, bonds
in the wall will rupture. How small a radius and how thick a wall
can we choose before the tube becomes unstable? Even if a tube
does not spontaneously break, how much external force can it
tolerate? It's quite straightforward to analyze these issues
using the molecular mechanics techniques in PolyGraf, a molecular
modeling software package produced by Molecular Simulations,
Inc.. These methods include energy minimization using any one of
several empirically derived forcefields (such as MM2), molecular
dynamics, and analysis of strain by examination of bond length
and bond angle data from the minimized structure, etc.
Given two tubes of differing radii, we can insert the tube of
smaller radius into the tube of larger radius. This creates a
simple bearing, for the outer tube can rotate with respect to the
inner tube. What surface is best in such a bearing: the 001, the
110, the 111? Some other? With what orientation? How different
should the radii be? If we grip the two tubes and try to pull
them apart, how much force must we apply? Again, using molecular
mechanics it is not too difficult to get reasonable answers to
these questions.
Figure 1. A 2,808-atom steric-contact sleeve
bearing, based on concentric cylinders of strained
diamond with modified (100) surfaces (oxygen and sulfur
termination). Energy barriers to rotation are small
compared to thermal vibrational energies. Designed by E.
Drexler at Xerox PARC using Polygraf software and a
computer-aided design tool under development by R.
Merkle.
Figure 2. Same as Fig. 1, in an exploded view.
Interlocking ridges on the shaft and sleeve give large
strength and stiffness in resisting axial displacements
at the sliding interface.
If the tube is being used as a component in a positioning device
(perhaps a nano-arm for an assembler), we would like to know its
stiffness. How far will it bend under the influence of thermal
noise? What changes can we make to improve the stiffness? If we
push on something with our nano-arm, how much will it bend?
While tubular structures are quite interesting, they are only one
of many components in an assembler, and this program is only the
first of many other similar programs. Of course, at some point
we'll want to design a language that can be used to describe the
different kinds of nano-components, and a compiler to turn
descriptions in that language into actual atomic coordinates, but
that's a future project.
Lakshmikantan Balasubramaniam (usually known as Bala, it's much
easier to pronounce!) has joined us at PARC for the summer to
work on these and other problems that arise in molecular
manufacturing. Bala is a graduate student in the mechanical
engineering department at MIT specializing in computer aided
mechanical design.
Dr. Merkle's interests range from neurophysiology to computer
security; he heads the new Computational Nanotechnology Project
at Xerox Palo Alto
Research Center.
Has
Penrose Disproved AI?
One of the most talked about and reviewed books of recent
years is Roger Penrose's The Emperor's New Mind
(1989, Oxford Univ. Press). So why publish yet another review?
Because the popularity of a book whose jacket declares it
"dares to suggest that the emperors of strong AI have no
clothes" has apparently given some casual observers the
impression that Penrose has dealt a death-blow to artificial
intelligence (AI). This is not even close to being right.
Being read is not the same as being believed. Most reviewers have
praised the book as original, well-written, thought-provoking,
etc., and then gone on to take issue with one or more of
Penrose's main theses.
Penrose seems unfamiliar with the existing literature in
cognitive science, philosophy of mind, and AI. The handful of
reviewers who agree with Penrose don't seem to have paid much
attention to his specific arguments--they always thought AI was
bogus. See, for example, the 37 reviews in Behavioral and
Brain Sciences (BBS), Dec. 1990, V13,
pp.643-705.
But aren't most of these reviewers fuzzy-headed philosophers of
mind and computer science researchers, while Penrose is a good,
solid, world-renowned mathematical physicist? Penrose himself
repeatedly emphasizes the speculative nature of his musings,
warning that "my point of view is an unconventional one
among physicists and is consequently one which is unlikely to be
adopted, at present, by computer scientists or
physiologists."
But if appeals to consensus and authority won't persuade you,
let's get down to details. First let me agree with most reviewers
that this is a great book. It makes the reader think. Most of the
middle of the book is a wonderful tutorial on various subjects,
mostly in physics. If you understand Penrose's discussions of
entropy, for example, you can easily see why cosmological
theories of "inflation" cannot deliver what their
proponents claim. Wrapped around these tutorials, however, and
mostly confined to the introduction and conclusion, is a sloppier
collection of arguments for what is clearly a deeply-felt
opinion: "Yet beneath this technicality is the feeling that
it is indeed 'obvious' that the conscious mind cannot
work like a computer, even though much of what is actually
involved in mental activity might do so. This is the kind of
obviousness that a child can see..."
Penrose grants that we may be able to artificially construct
conscious intelligences, and "such objects could succeed in actually
superseding human beings." But he thinks "algorithmic
computers are doomed to subservience."
Penrose's argument is two-fold. First he tries to show why
human-type intelligence could not be implemented by any
Turing-machine equivalent computer (ordinary, parallel, neural,
or otherwise). Then he tries to show how it could be physically
possible that the human mind is not algorithmic in this sense.
Penrose gives many reasons why he is uncomfortable with
computer-based AI. He is concerned about "the 'paradox' of
teleportation" whereby copies could be made of people, and
thinks "that Searle's [Chinese-Room] argument has
considerable force to it, even if it is not altogether
conclusive." He also finds it "very difficult to
believe ... some kind of natural selection process being
effective for producing [even] approximately valid
algorithms" since "the slightest 'mutation' of an
algorithm ... would tend to render it totally useless."
These are familiar objections that have been answered quite
adequately, in my opinion. But the anti-AI argument that stands
out to Penrose as "as blatant a reductio ad absurdum
as we can hope to achieve, short of an actual mathematical
proof!" turns out be a variation on John Lucas's
much-criticized "Gödel" argument, offered in 1961. A
mathematician often makes judgments about what mathematical
statements are true. If he or she is not more powerful than a
computer, then in principle one could write a (very complex)
computer program that exactly duplicated his or her behavior. But
any program that infers mathematical statements can infer no more
than can be proved within an equivalent formal system of
mathematical axioms and rules of inference, and by a famous
result of Gödel, there is at least one true statement that such
an axiom system cannot prove to be true. "Nevertheless we
can (in principle) see that P_k(k) is actually true!
This would seem to provide him with a contradiction,
since he ought to be able to see that also." This
argument won't fly if the set of axioms to which the human
mathematician is formally equivalent is too complex for the human
to understand. So Penrose claims that can't be because "this
flies in the face of what mathematics is all about! ... each step
[in a math proof] can be reduced to something simple and obvious
... when we comprehend them [proofs], their truth is clear and
agreed by all."
And to reviewers' criticisms that mathematicians are better
described as approximate and heuristic algorithms, Penrose
responds (in BBS) that this won't explain the fact
that "the mathematical community as a whole makes
extraordinarily few" mistakes.
These are amazing claims, which Penrose hardly bothers to defend.
Reviewers knowledgeable about Gödel's work, however, have simply
pointed out that an axiom system can infer that if
its axioms are self-consistent, then its Gödel sentence
is true. An axiom system just can't determine its own
self-consistency. But then neither can human mathematicians know
whether the axioms they explicitly favor (much less the axioms
they are formally equivalent to) are self-consistent. Cantor and
Frege's proposed axioms of set theory turned out to be
inconsistent, and this sort of thing will undoubtedly happen
again. Apparently, Penrose didn't do his homework on the Gödel
issue.
Penrose raises one issue that I do think deserves closer
scrutiny, namely exactly what sort of "motion" an
algorithm would have to be put into before it could subjectively
"feel" and be conscious. If we wrote down an algorithm
equivalent to Einstein in a book, "would the book-Einstein
remain completely self-aware even if it were never examined or
disturbed by anyone?" This issue has been raised before, and
is not particularly threatening to computer-based AI, but is
interesting nonetheless.
The other half of Penrose's arguments is a speculative "germ
of an idea" about how it could be that people are devices
which can compute things that a Turing machine can't, even though
all known physical laws do not allow the construction of such
devices, and even though "Most physicists would claim that
the fundamental laws operative at the scale of a human brain are
indeed all perfectly well known." One usually describes the
evolution of a quantum system by two processes, U and R, acting
on quantum states. Usually the unitary process U is in control,
but on occasion (when, exactly, is not very well understood) a
reduction process R intervenes.
Penrose is (refreshingly) a firm realist about quantum mechanics,
believing both these processes and the states they act on are
quite real and independent of observers. Penrose speculates that
this view will have to corrected somewhat when quantum mechanics
is integrated with general relativity. Penrose hopes that a
quantum gravity R will be exactly non-deterministic enough to
counterbalance the merging of state trajectories due to the
evaporation of black holes, and just time-asymmetric enough to
satisfy his thermodynamics-explaining conjecture that the Weyl
curvature approaches zero at past singularities. This R will
happen when the difference between components of a quantum
superposition approaches a one virtual graviton level, so as to
avoid awkward superpositions of differently shaped space-times.
We will need a radically new concept of space-time to deal with
simultaneity problems of an objective reduction law. Oh, and one
other thing: At the "borderline which interpolates between U
and R" "some new procedure takes over." "This
new procedure would contain an essentially non-algorithmic
element" so that "the future would not be
computable from the present, even though it might
be determined by it."
Now, general relativity doesn't affect life on Earth much, and
its effects are very hard to discern even in astrophysical
contexts. Quantum gravity should be a small correction to general
relativity, revealing itself in even more unusual circumstances,
such as distance scales of 10-35 cm. Nevertheless,
Penrose speculates that this new quantum gravity U/R
interpolation procedure is how nature assembles the recently
discovered quasi-crystals since "the general tiling problem
... is one without an algorithmic solution" because of
non-local constraints.
Similarly "somewhere deep in the brain, [as yet unknown]
cells are to be found of single quantum sensitivity" so that
"synapses becoming activated or de-activated through the
growth or contraction of dendritic spines ... could be governed
by something like the processes involved in quasicrystal
growth," simultaneously trying out "vast numbers [of
possible alternative arrangements], all superposed in complex
linear superposition." All this somehow affects only our
conscious mind, leaving our unconscious to compute
algorithmically, and quantum non-locality explains "the
'oneness' of consciousness." "True intelligence
requires consciousness" and the conscious mind (of
mathematicians) has "a direct route to truth, the
consciousness of each being in a position to perceive
mathematical truths directly."
Penrose believes "mathematical ideas have an existence of
their own, and inhabit an ideal Platonic world, which is
accessible via the intellect only." "The mind is always
capable of this direct contact. But only a little may come
through at a time." This contact is what explains "the
deep underlying reason for the accord between mathematics and
physics." I am not making this up. If you are not familiar
with modern physics or physiology, I do not know how to convey to
you just how unlikely Penrose's scenario is, except to offer 100
to 1 odds against it. Yes, it is logically possible, but only if
everything goes just Penrose's way.
A more popular quantum realist position than Penrose's (though
quantum realists are still a minority) is the
"many-worlds" view, which says there is only the
process U and that R is an illusion. (I'd bet this has at least a
1 in 20 chance of being the closest we have to right.) Penrose
rejects this view because "a theory of consciousness would
be needed before the many-worlds view can be squared with what
one actually observes." That is, we don't know how to test
it yet.
In BBS, Penrose expressed surprise at how many AI
reviewers support the many-worlds theory, and makes a snide
comment about trusting "that their reasons for believing in
the validity of the AI programme are more soundly based."
Yet the review in Science by a physicist also
supported many-worlds. Moreover, many-worlds is especially
popular among quantum gravity researchers, and one of Penrose's
main plausibility arguments comes from a demonstration by Deutsch
that with many-worlds a 'quantum computer,' though still
algorithmic, could get a large speed-up relative to an ordinary
computer.
Martin Gardner calls Penrose's book "the most powerful
attack yet written on strong AI." If so, AI must be doing
pretty well. If the book were condensed to a paper by deleting
the excellent tutorials, and if Penrose's name weren't on it, I
doubt if the paper would have been much noticed, or even
published.
Regardless of your opinions about the appropriateness of current
AI research strategies, or about the length of the road ahead,
Penrose's book offers no substantial reasons to change your views
about the long-term possibility of computer-based AI. The fact
that many casual observers have been misled about this is yet
another indication of the inadequacy of our current methods of
forming and communicating scientific consensus.
Robin Hanson (hanson@charon.arc.nasa) does AI research at NASA
Ames, has degrees in physics and philosophy of science, and on
the side studies alternative methods for scientific consensus.
Ecotech
Conference
The prospects for using molecular manufacturing techniques to
benefit the environment will be explored at the Ecotech
conference, a nonprofit project of the Tides Foundation to be
held November 14-17 in Monterey, California. Eric Drexler will
speak on how molecular technologies could both reduce
environmental damage from new production and clean up
already-existing wastes. Following the talk the Foresight
Institute will hold a workshop to explore participants' ideas on
what can and should be done. In addition, Foresight will be on
hand in the "marketplace of ideas" section of the
conference to discuss how nanotechnology-based tools and products
can aid environmental restoration. In addition to educating
attendees, we hope to build broader contacts with experienced
environmental activists sharing an interest in the potential of
molecular manufacturing.
Nanotechnology is just a part of this large meeting devoted to
encouraging companies to make environmental issues an integral
part of their overall business strategies. The meeting will help
managers, strategists, planners, communicators, educators,
lawyers, legislators, investors, scientists, and entrepreneurs
understand how they can participate in solving the tough
challenges that lie ahead.
An intriguing mix of speakers has been lined up, some of whom you
may recognize as friends of Foresight. A few of the better-known
speakers are entrepreneur/author Paul Hawken; Peter Schwartz,
president of Global Business Network (which cosponsored the First
Foresight Conference on Nanotechnology); Paul Saffo, research
fellow at the Institute of the Future; Whole Earth Ecolog
editor J. Baldwin; and scientist/author Amory Lovins.
Conference attendance is limited to 450 registrants. Registration
fees are $500 per participant and $250 for representatives of
nonprofit organizations. For more information, contact Mike
Whitacre at 619-259-5110.
From Foresight Update 12, originally
published 1 August 1991.
Foresight thanks Dave Kilbridge for converting Update 12 to
html for this web page.
|