Wolfram and Kurzweil Roundtable Discussion
"The most dramatic possibility is the universe started from a simple initial condition that had some simple geometrical symmetry. It might be the case that if we turn our telescope off to the west, and look at the configuration of the universe in the west, it might be identical to the configuration of the universe in the east [...]"
Originally recorded for BusinessWeek.com
July 20, 2005. Transcribed and reprinted on KurzweilAI.net February
24, 2006.
Otis Port: Welcome, we're here to examine the implications of a
controversial new book, A New Kind of Science by Stephen
Wolfram. We believe the complex systems have complex causes and
therefore the way to comprehend such a complex thing is to carve
it up into little pieces and study each one of those individually
until we understand the piece well enough to express it as a mathematical
formula. Stephen says not so. Complex systems don't have complex
causes. In fact, they have very simple mathematical type algorithms
at their roots. And we can simulate, emulate nature's algorithms
with something called cellular automatons or cellular automata,
CA for short. He believes that eventually it will explain everything,
literally everything from what happened; how the cosmos evolved
since the Big Bang 15 billion years ago to the evolution of all
biological life and intelligence. He hasn't yet found the mother
of all CA's that one master cellular automaton that would explain
everything that has happened since the Big Bang, but he's certain
that it exists, and he's hopeful that he'll find it and he's certain
that it will be astonishingly simple when he does.
Joining us also is Ray Kurzweil. He's a computer scientist,
inventor, author, and CEO of Kurzweil Technologies and a couple
of other related companies. The thread running through most of Ray's
work is Artificial Intelligence or AI. Computer Automata are a common,
have been a common tool for AI researchers since the 1970s so when
Stephen's book came out we asked him grab it and critique it
for us. And there's a full
version of his critique on KurzweilAI.net.
Basically, Ray was fascinated by most of Stephen's work and
impressed with its breadth, but he quibbles that computer automata
can really evolve into the complex systems that Stephen believes
they can.
My name is Otis Port. I write about technology, and AI has been
something that has intrigued me for ages. Gentlemen, thank you both
for being here, Stephen, maybe a place to start would be the one
cellular automata that gets a lot of attention in the book, either
30 or 110, and something like that and why it's special.
Stephen Wolfram: I got interested a long time ago twenty years
ago, now, in the question of what the right primitives to understand
the natural world should be. There's been a tradition for three
hundred or so years of using mathematical equations, the kind of
constructs of human mathematics as the appropriate primitives to
describe the natural worlds, but with computers one has a way of
thinking about those things in terms of programs. I got interested
in the question of what would happen if one would allow programs
that generalize things like mathematics. So, I was interested in
the question of well, what does simple programs typically do. Well,
our usual intuition, my intuition was if the program is sufficiently
simple, then what it does must be correspondingly simple. That's
the intuition we get from for example our usual experience with
engineering. If we want to make something complicated we expect
that we have to go to a lot of effort and make very complicated
rules and set things up in a very complicated way to achieve complicated
results, but I decided to do a very systematic computer experiment
and try all the possible simple programs of a particular kind. When
I went through and looked at these programs, they can be numbered
in some straightforward way. When I got to rule number 30 I saw
this amazing thing. I had made a pattern on a computer display,
actually it was a printout at the time. One started out from just
one black cell at the top followed this very, very simple rule and
when one traced it along, the pattern that it made was incredibly
complicated and this to me was very amazing. At first I assumed
that the fact that it looked complicated must be a feature of sort
of inadequacies of my visual system that really there was regularity
but I just couldn't see it. So, I tried all sorts of elaborate
mathematical and statistical tests and so far as any of them could
tell this thing was complicated. Well, that observation kind of
ended up changing my whole view of many of the foundations of science.
It's something that took me years to come to terms with.
What it revealed was this point that basically if you pick that
even with very simple programs and even very simple underlying rules,
it's possible to get very complicated behavior. Once one's
found something like this it seems very obvious seems like how could
this not have been found a zillion years ago. How could I not have
found it several years earlier than I did? Now, having discovered
that very simple rules can do complicated and interesting things
you can go back and look historically was this seen before. The
answer is absolutely yes. If you go back to ancient Greek times
for example and look at the Pythagoreans talking about prime's
for example, prime numbers. There's a fairly simple rule for
generating prime numbers. Yet, once generated the primes make a
kind of complicated and hard to predict sequence. That's really
an example of the same kind of phenomenon I ended up finding in
rule 30 and so on and the emphasis for example in studying primes
was to find regularities in the distribution of primes not focusing
on this phenomenon of even though the rules for primes are simple
the actual form of primes can be complicated. If that point had
been recognized in antiquity, then my guess is that a lot of kind
of the direction of how science has developed would have been somewhat
different. People would not have been as mystified as how can nature
be as complicated as it is. I'm not sure if Ray agrees with
this. But perhaps we should find out.
Ray Kurzweil: First of all, I don't doubt the possibility
that we could explain the whole universe as a cellular automata.
The question is how far does the concept of a cellular automata
get us? I did, by the way find the book quite delightful and I got
a lot of insights in reading it. There are a couple of issues. One
is, you describe how a simple cellular automata, rule 30 can create
designs that are incredibly complicated and it is pretty amazing.
You really would think that starting with one cell you would get
something really simple.
OP: Let's tell people who don't know how a cellular automata
evolves how it does it. The fact that you start with…
RK: A cellular automata, there's different ways of doing it.
The simplest one is you have one cell and then the next layer of
cells will be dependent on that cell and adjacent cells. And it
is pretty remarkable to look at them and you have an assemblage
of features even some nesting of features. Triangles are different
sizes, and lines and streaks and it seems to have a mind of its
own and you think it's settling in but then something else
happens and it's really unpredictable, but it's more than
just unpredictable, because pure randomness isn't very interesting
either. Pure randomness becomes predictable by its lack of predictability.
Here there seems to be order but then it lapses back into disorder
and it's interesting, patterns and pretty amazing that all
of this complexity comes from very simple rules and simplest possible
starting position of one black cell. The question though is how
complicated? Are they really incredibly complicated? Are there different
levels of complexity? I want to come back to that. I think it's
probably a key issue that we should talk about and whether or not
for example the complexity and intelligence of the human brain is
perhaps the one entity that we can point to that is as complicated
our intuition is that it's as complicated as anything else
we know about and we're trying to emulate that in our computer
science, and it's the whole field of Artificial Intelligence.
How does that compare, say to the complexity of a dust storm or
fluid turbulence and things like that. I think that's really
a key issue.
To start out with a lesser but also interesting issue. The issue
of whether or not the universe ultimately cellular, that is to say
fully digital or analog and Stephen shows how you can take these
certain simple rules and produce different types of patterns. For
example the pattern on a tent olive shell which is pretty interesting
and satisfying and aesthetically pleasing pattern. Actually, it
does resolve very simply from a cellular automata. So, these simple
digital rules can provide certain types of patterns.
But it's also possible for analog phenomena to produce similar
phenomena and Stephen you actually provide that derivation in the
book. Showing how analog continuous functions also can provide the
same types of patterns, so if we see these types of patterns in
nature it does not necessarily imply the inverse conclusion that
it necessarily was caused by a digital cellular pattern. It could
have been produced by something analog, and—
OP: Stephen, you're sure that the universe is digital?
SW: Well, so one thing to understand when we're talking about
for example mollusk shells and we're talking about what makes
the patterns on those shells and so on, you're talking about
making a model that is in an a sense an approximation, an idealization
of the way an actual mollusk work, and we're not expecting
to reproduce in every precise detail sort of the whole digestive
process of the mollusk and so on. There is one case in natural science
where modeling works differently. And that's if you can actually
get an ultimate fundamental model for the universe, because that's
a case where the model is no longer an idealization, no longer an
approximation, if there really is an ultimate model for the universe
it's just this is precisely how the universe works, and at
that level you can very sharply ask the question is it a model that's
based on discrete elements is it a model that's based on continuous
kinds of numbers of the type that have been studied in calculus
and so on.
RW: Let me comment on that. I think it's an intriguing notion
and such a model would be a great step forward and Stephen takes
some significant steps towards creating that type of model, in the
book. There's more work to be done and I understand that you're
continuing to work on developing a theory of physics that would
take into account all the different things we know from the standard
model and so on.
Let me read you something from my book, from my first book The
Age of Intelligent Machines. I discuss the question whether
the ultimate nature of reality is analog or digital. I point out
that as we delve deeper and deeper into both natural and artificial
processes we find that the nature of process often alternates back
and forth between analog and digital representations of information.
I noted how the phenomena of sound flips back and forth between
digital and analog representations. In our brain music is represented
as a digital firing of neurons in the cochlear representing different
frequency bands in the air and in the wires. The music in loud speakers
is an analog phenomena. Representation of sound in a music compact
disk is digital, which is interpreted by digital circuits, but the
digital circuits consisted of threshold transistors, which are analog
amplifiers. As amplifiers the transistors manipulate individual
electrons which can be counted and therefore digital, but on a deeper
level an analog subject to quantum field equations, and at a yet
deeper level Stephen Wolfram and some other theorists have theorized
that digital computation basis to the continual equations and along
those lines that's a question, suppose we find a computer at
the basis of the universe that's digital and follows these
rules, it would be a great accomplishment to represent everything
we know in a cellular automata type framework but it wouldn't
necessarily be the final answer. We wouldn't know that there
are deeper analog phenomena.
SW: Several things to say. First of all, these kind of progressions
of different kind of descriptions of things in analog form or digital
form, this is you have to remember that all these kind of descriptions
are always approximations. They're always idealizations. When
we have the digital representation of a sound on a CD for example,
it's the idealization of the sound it's not representing
how every molecule in the vibrating air works; it's some kind
of idealization. The thing that is different and very shocking about
an ultimate theory of physics, it's no longer an idealization.
It would be if I'm right there's a theory, there's
a set of rules, which just is the universe. If you say how are these
rules implemented, that question doesn't really make that much
sense because—
RK: Our experience in physics is that we have discovered finer
and finer models. We have these discrete electrons. They just are.
But then we find there are more complex field equations underlying
them and they're [inaudible] and we find if string theory is
correct there are actually pretty cellular sounding type mechanisms
underlying those field equations.
SW: The most dramatic possibility is the universe started from
a simple initial condition that had some simple geometrical symmetry.
It might be the case that if we turn our telescope off to the west,
and look at the configuration of the universe in the west, it might
be identical to the configuration of the universe in the east. That
would be the most bizarre possibility if the thing started from
some simple initial condition and we get to see enough of the universe
that we can actually sort of tell the kind of progeny of one side
of the initial condition and the other side of the initial condition.
In most theories that have inflationary things happening we don't
get to see enough of the universe to actually be able tell that.
OP: I'm sure this is a silly question, Stephen, but if you
actually find that and we actually run it on a computer aren't
we creating another universe?
SW: It's not a universe. It's just an emulation of a
universe.
OP: But you said, didn't you say a little while ago that if
you found that digital code it would no longer be an idealization
it would be.
SW: That's correct. So, if we were to run it for long enough
on a computer, we would be able to find out what the universe would
do in precise detail.
RK: But we don't have enough time to run it. For most cellular
automata given the fact that all these different phenomena interacting
with each other, you cannot predict the final outcome without literally
simulating every step and since we don't have computers that
run faster than the universe itself at it's finest level of
granularity we couldn't really run a model of the universe.
We could run the early part and have some idea of how it started,
but it would never catch up to us because it can only run as fast
as the universe runs or slower, generally speaking.
SW: I want to come back to the question of sort of the intuition
that one has based on the history of physics, and how that relates
to what one might now believe is true. It's certainly the case
that if one looks at the history of physics, that every time one
gets to a greater level of smallness and analysis of the physical
world, things have gotten more complicated, there's another
layer and so on.
One of the things that in a sense has been the case is the kind
of the primitives that have been used in understanding those layers
are typically mathematical primitives and I think they've been
somewhat constrained by people's interest in having primitives
where one can kind of foresee what the outcome of doing particular
things with those primitives will be. I think one of those things
that has come out from this sort of experimental work I've
done on what simple computer programs do is that with more general
kinds of primitives, much more surprising things can happen. One
can really get much greater richness from much simpler ingredients.
I think that's what changed my intuition about whether there
could really be an end to this sort of sequence from classical mechanisms,
quantum mechanisms, quantum field theory, string theory. Whatever
else. I mean, I might also say in terms of the continuous versus
discrete dichotomy, when things get as abstract as they have to
be if they really is a very simple a model for the universe, and
I mean, one has to realize that if one can describe a whole universe
in terms of rules that are very short, then it's sort of inevitable
that nothing that's familiar to us will be obviously visible
in those rules. There isn't room to fit in a three for the
three dimensions of space or the mass of the electron or something
like that so what has to be there in these underlying rules must
be quite abstract and quite unfamiliar to us.
RK: But, then we have to question is there a deeper theory. And
I'm not sure how you would ever prove there aren't in
fact some more deeper phenomena.
SW: I think the way you would do that as a practical matter. If
it were the case that the simple rules that one knew reproduced
everything we know in our universe, if one could establish that
with reasonable certainty then what would be the point of saying
there's something deeper. It would have no relevance to what
would actually happens in our universe.
RK: Let me bring up a key issue one where I think I have a somewhat
different perspective than you expressed, Stephen, which has to
do with levels of complexity. You said that the images from a class
four automata were incredibly complicated and in the book there
are a lot of statements along the following lines “what I come
to believe is that many of the most obvious examples of complexity
in biological systems actually have very little to do with adaptation,
or natural selection but instead they are mainly just another consequence
of the very basic phenomena that I discovered that along with any
kind of system, any choices of underlying rules inevitably lead
to behavior of great complexity and I have some issue with the concept
of great complexity. I think there's something missing in between
a CA and what it can do by itself and the kind of phenomenon in
the world I would consider of great complexity like human beings
or Chopin preludes or insects, things like that. If we run a class
4 automata, it has some features and they're unpredictable
and there's a certain amount of order, but you could run it
for a million or a trillion or trillion trillion iterations and
you still get the same kind of level, same kind of pattern of these
features and you don't have emerging anything comparable to
insects, let alone humans. So, in the conversation we had before
we kind of tossed around is this a concept of intelligence, which
I would put forth as a phenomena that is an order of complexity
greater than what we see in a class 4 automata, fundamentally greater.
Is there such a thing as that concept and you said, well, intelligence
is really culturally determined, and so I went looking at various
definitions and indeed the vast majority of them are culturally
determined, but I think there is a way of defining intelligence
that strips away human specific characteristics.
SW: I'll be interested to hear this.
RK: And I'll read a couple of things. Intelligence represents
a hierarchy of features that solves unexpected problems better more
quickly, better results than a lesser hierarchy. There's a
nesting of features now you get a little bit of nesting of features
in a class 4 automata, but it's a pretty simple level and it
never really gets to be a higher hierarchy. We have a very elaborate
hierarchy of features within a human being and certainly if you
go to a very fine level, you'll see a lot of chaotic chaos
and the amount of computation in a human could be the same as the
amount of computation in a dust storm or even in a rock at a fundamentally
detailed enough level, but if we look at it in the right level,
we see a very elaborate hierarchy, and it's not an arbitrary
hierarchy. Each level has actually contributes something in terms
of being able to solve problems and it involves the ability to create
models of the world. I know you have some simple examples of CA's
or CA like processes that create simple models, but the models that
we create are themselves hierarchical and can be as hierarchical
as even the human brain that created it. I've got definition
of intelligence here in The Age of Spiritual Machines, the
ability to use optimally limited resources, including time to achieve
a set of goals which may include survival, communications, solving
problems, recognizing patterns, performing skills the faculty of
[inaudible] which orders perceived in a situation previously considered
disordered. That's a quote from R.W. Young. And then the concept
of order, information that fits a purpose and the measure of order
is how well the information fits the purpose. In an evolutionary
algorithm, the purpose is to solve a problem. I'm not saying
that this shows that the world is not a CA, but I'm perfectly
willing to accept that possibility, but you need something on the
substrate of the CA to get to these higher levels of order and in
my view that is precisely an evolutionary process some kind of selection.
Evolution is needed. Cellular automata alone doesn't get you
there.
SW: One feature of us as humans and I am no exception is that we
would like to think that we're special. This has been something
that has been true throughout the history of science, human thinking
and so on. We'd like to think that there are ways in which
we are not only in detail special but in generality and abstractly
special. And that's something that history of science has not
been kind to the idea that humans are special. I mean for example
four hundred years ago we find out that the earth isn't at
the center of the universe. A hundred and fifty years ago we find
out there isn't anything special about the origin of our species.
RK: But there I something special about humans at least if we ask
the questions of how are we different from different species. We're
the only species that has ushered in its own evolutionary process
which is technology. You can find other animals that use tools,
but those tools do not embody a knowledge base that gets passed
on from generation to generation. With the knowledge base itself
expands and grows exponentially over time. We're the only species
that expands our own horizons.
SW: I'm not sure. We don't know with whale songs for
example, we don't know how exactly those are passed down. With
bird songs there's a certain degree of passing them down, and
I think that obviously.
RK: They don't have a technology. There's a limit to
the whale songs' complexity because they don't have means
of writing it down. They don't have technology. There's
no indication that it's significantly evolving. Whales did
not go into the air. They didn't go off the planet. They're
not extending their life spans. They don't have video cameras;
they don't publish books on cellular automata. They don't
debate issues about whether or not they're special.
SW: We don't know that. I think that this issue, you mention
lots of details and one can argue to us they're very important.
The way civilization works, the way technology works, these are
things that are very important to our specific lives, but—
RK: It's specifically an evolutionary process. Our combination
of pattern recognition, cognitive function and our ability to manipulate
the environment through our opposable thumb has allowed an evolutionary
process to continue in the guise of our technology and it evolves.
SW: One of the things as we get to things like the opposable thumb,
the opposable thumb is surely a detail. We can't hang our theory
of intelligence on the opposable thumb. That would be a very peculiar
circumstance if the intelligence and the consequences of intelligence
were critically dependant on the opposable thumb.
RK: I think our technology is dependant on ability to take models
and I don't doubt that even birds can model their environment
to a certain extent and clearly giant squids are very intelligent.
Whales. And they clearly have some model of their environment but
the opposable thumb allowed us to take our models and build them
and then continue to improve them exponentially to the point where
they're actually expanding our own horizons. We've more
than doubled our life span in the last hundred and fifty years.
We went off the planet. We're going to reverse engineer our
own brains and understand how they work and build non-biological
analogs and none of that could happen unless we had the ability
to take those models in our brains and build them in the real world.
SW: I think one of the issues is what is the abstract essence.
We know that there are certain details of human intelligence, some
of the things you've mentioned that are significant to us quite
special as details. The question is there's this abstract version
of intelligence where we can say that something like a turbulent
fluid doesn't have. We know it doesn't have the details
of human intelligence.
RK: It doesn't have the hierarchy. There's a little bit
of hierarchy in that you have atoms and molecules and there may
be little eddies and so on that are features but it's limited.
It does not have the level of complexity, the order of hierarchy
that we see in a complex system.
SW: I'm being unfair, but in turbulent fluids are a bad example
for not having hierarchy because they have this Kolmogorov cascade
and eddy sizes and so on.
RK: I'm not saying no hierarchy. I'm saying they don't
have as elaborate a hierarchy as.
SW: The issue is, what I thought when I was working on my book,
what I thought when, I was quite sure because I'm as egoistical
as the rest of us but there was something very special about humans
and about our intelligence and the kinds of things that I was studying
that were relevant to physical science and natural science and so
on would not reach that. That was the prejudice that I had going
into what I did. As I investigated more and more I became less and
less convinced of that prejudice until I got to the point where
I just don't believe it at all. And that there were several
steps that got me to that point, but maybe you wanted a more—
RK: I agree and I disagree. I disagree that there isn't something
special, but I think that the specialness is a very abstract one,
specifically that we expand our horizons and other species don't.
We're kind of the vanguard of evolution in that sense, but
I agree that there have been failed attempts to define our specialness
some as simple as the fact that all the stars are obviously going
around the earth and a lot of different conceptions that failed
to hold water, but you seem to take that to the extreme of saying
there is no such thing as intelligence, different orders of complexity.
The doctrine of computational equivalence basically says that the
kind of computation we see in the brain is equivalent to that in
a dust storm or fluid turbulence and at a certain level I agree.
They are both computational processes, they can both run on any
kind of universal computational medium. Computation is simple and
ubiquitous, and I agree with all that and intelligence is extremely
powerful concept and it emerges from evolution and the exact form
of any form of intelligence is very arbitrary. Human beings are
the way we manifest our intelligence is through an extraordinary
accident of our vine of evolution and any intelligence that may
evolve would go through some intricate and fairly arbitrary chain
and would end up being culturally determined within it's own
culture and none the less you can strip away the culture and it
does have a necessary hierarchy of intricate features that can solve
problems and perform tasks.
SW: I would have loved to conclude the kind of things that you're
saying. I would have liked to find sort of the essence of intelligence.
The thing that what makes us different from these other things that
I believe have extensive computational abilities. I have not managed
to find that. Now, if somebody can produce it I will become a very
happy person, but I have not managed to find it. So the question
is does the weather have a mind of its own. And that's where
you would say it does not and I would say it's very hard to
pin down what you mean by that.
OP: Ray brought up; he mentioned your theory of computational equivalence
a little bit ago. Help me out a little bit. How can you be sure
that well, first of all, you postulate right that once complexity
reaches a certain level it doesn't get any more complex? How
do you know that?
SW: In science, most of the time, one gets to make inferences based
on ones observations and so on, so the principle of computational
equivalence is something that's a merged from lots of experiments
and analysis that I've tried to do. Inevitably in science it's
a leap beyond what I can know from my experiments and my observations.
OP: Have you actually tried to make something more complex and
couldn't?
SW: No. That's not the way this tends to work.
OP: Okay.
SW: The idea of the principle of computational equivalence is essentially
any system where its behavior is not sort of obviously simple and
regular will tend to correspond to a computation that is at the
same level. Now, it's not clear that there might not be things
that are a higher level for example people as a theoretical matter
even when Alan Turing first wrote about universal computers and
turing machines and so on, he said well, what if you had something
that was beyond a turing machine that was a turing machine plus
what he called an oracle which is a black box that answers questions
a turing machine can't answer. Then at least, in principle
we would have something that was beyond a turing machine that was
able to do more computation than a turing machine and cellular automaton,
and any of the other kinds of programs I looked at. Then it's
a non-trivial statement to say no, those kinds of oracle black boxes
while in principle we can talk about them don't actually exist
in our universe. You can't actually make one of those in our
universe. Actually, let's me say another thing though. It's
when we're talking before you were observing that one of the
challenges that you've often faced is people saying AI isn't
possible, the brain, what humans do is just something beyond what
we could ever achieve with technology, with machines and so on.
And I think that to many people perhaps, still there's some
doubt, can machines replicate all of the fine intuition, emotions,
language, whatever other properties of the brain they think they're
particularly concentrating on. I think you and I agree, and probably
disagree with a great many other people that really it will be possible
to replicate all of the features of a brain in a piece of solid
state electronics or something.
Now, in a sense, what I'm saying and the conclusions I have
come to and are embodied in these principles of computational equivalence
and the things I said in the book are even more extreme than that.
Saying not only can we replicate what goes on is there a replication
of what goes on in the brain that could be achieved by solid state
electronics but that sort of this is the kinds of things that go
on in the brain are naturally reproduced by… there's no
essence of what goes on in the brain, that isn't already something
that happens in lots of these natural systems that show behavior
that seems as complex.
RK: Clearly there's a hierarchy of features in the brain.
It's able to create models of itself and of the real world.
Those models have hierarchies like language has a lot of intricate
features in it and language is a product of the brain, how do you
get from rule 110 to human brain and I'm not saying you can't
get there, but there's only two ways. One is through some evolutionary
process, and if you follow that elaborate evolutionary process,
you'll get to something culture determines is very different
than the brains. Or, you can reverse engineer some entity that's
already gone through the process, which is the AI model that I've
used which is still derivative of that evolutionary process.
SW: One thing that I think is in a sense ironic. I'm sure
you run into many people who say you know how can you possibly make
a hardware AI? How can you possibly reproduce all those details
of the brain and so on and so on and so on? What I would say in
response to that is you'll be arguing about that until it really
exists.
RK: That's right.
SW: There's no way to abstractly say whether we can do it
or not, and I'm going to have to say the same thing to you
about the extent to which one can produce something that has all
the features of our intelligence.
RK: Without going through an evolutionary process and without copying—
SW: Let's take a couple of thought experiments that I think
are perhaps interesting. One thing I think a useful collection of
thought experiments come from trying to think about extraterrestrial
intelligence because extraterrestrial intelligence is another intelligence
that doesn't share the historical details of our intelligence
so the question is how would we tell whether we're seeing an
example of extraterrestrial intelligence. It's interesting
to look at the history of what's happened in attempts to identify
extraterrestrial intelligence. I think it's sort of constructive.
For example, famous example, in the early days of radio, Marconi
had a yacht that he used to fly the Atlantic on, I believe, and
on his yacht he used to do experiments so he set up a radio antenna
on his yacht and he was just listening to what radio stuff is out
there in the cosmos, and he heard these funny kind of whoosh, funny
kind of sounds, and his immediate assumption was something as complicated
as that and with all the structure that that has must be the Martians
sending us radio signals. It turned out it was various physical
processes and the plasma and the ionosphere. Same kind of thing
happened with pulsars for example, the kind of regular pulses from
a pulsar, the immediate assumption was these were extraterrestrial
intelligence beacons. Turned out to be the details of it turned
out to be rotations of a neutron star.
RK: They also turned out not to be that complex to have that many
features to really represent anything that interesting. It didn't
give, say a series of axioms in any arbitrary system in you would
say the Euclidian geometry is very arbitrary but in any axiom system
and then develop theorem's from it or some kind of abstraction
that we would recognize as intelligence and we dismissed it because
it didn't have those that kind of level of complexity and it
had a little bit of complexity because I think that natural systems
like wind storms in the cosmos have a certain amount of order because
they have also evolved to a certain level, but not to the level
of human brains.
SW: So, let me ask you this then. We see a signal from the cosmos.
How do we know if it corresponds to an—If it was produced by
intelligence.
RK: It could perhaps tell some stories about how planets are organized
and go around the star and how satellites go around the planets
and sort of represent that in using some sort of language, per se
or some method of symbolism. I think mathematics is an example that's
used the most and I know you said a simple process could generate
axioms and theorems. I'm not sure that's true.
SW: The one feature of mathematics I, epistemology does not necessarily
the thing I was expecting we'd be covering here, but let's to one
issue about mathematics is to what extent is mathematics as we know
it a cultural artifact and to what extent is it a necessary feature
of abstract existence so to speak and I have come to a conclusion,
again, a conclusion that surprised me that mathematics as it is
actually practiced is a really very much a cultural artifact. It's
a long chain of historical development, which starts in ancient
Babylon where arithmetic was used for commerce, geometry was used
for land surveying and these things were progressively generalized
over the course of Mathematical....
RK: Numbers are pretty fundamental. I don't know that they're
that culturally determined and so axioms and interesting theorems
about number theory if you suddenly saw that coded in some fashion
that would be a pretty convincing demonstration of intelligence.
SW: See, number theory is a good example because what was very
exciting to the Pythagoreans was perfect numbers and so on has not
been that interesting to most of the history of mathematics. It's
tremendously hard to kind of see outside of the cultural box so
to speak which one is and—
RK: So how about simply a coding of prime numbers.
SW: It's a fairly simple process that generates primes.
RK: But we've never seen that in the natural world.
SW: Yes we have. For example if you look at the rings of Saturn
for example at the divisions in the rings of Saturn, the form of
those divisions depends on the relative primality of integers.
RK: It's just a few primes. If you actually counted the first
thousand primes it would be pretty impressive for a natural process.
SW: How many primes would you like it to count? I think probably
about ten you can tell in the rings of Saturn. Whether there's
some other case in astronomy where you can see a hundred I'm
not sure. I'd have to think about it for a bit to see where
one might see this. Again, that's a thin thread on which to
hang—
RK: You'd be impressed if you had a signal that just counted
primes or counted or just told a very nice elegant story of an axiom
and theorem system, we'd be hard pressed to discount that as
a natural process.
OP: One of the points you made in your review that caught my eye
was that Stephen's cellular automata don't seem to have
any competition any tension between them so there's no win
or lose type decision at various points.
SW: One of the things that had also surprised me. I had thought
that the kinds of studies I was making of simple programs and so
on would be kind of relevant to physics, and perhaps to some kinds
of and to other kinds of things like that but in biology there would
be a higher level of complexity, a higher level of processes going
no that would be associated with the whole Darwinian idea of adaptation
and natural selection and so on and so I was very surprised at an
empirical level it seems to me that I was never able to find those
things that sort of were higher level because there were constraints
where selection being applied.
RK: But you denied the reality of intelligence and kind of dismiss
it because it's hard to define and because there's controversy
about its definition and a lot of the definitions are culturally
determined, which I'm not so sure is a bad thing. Evolution
has accomplished something, and the difficulty in defining it and
a hierarchy in features is where not an arbitrary hierarchy as you
have in rivers and estuaries on just a planet, but a purposeful
hierarchy where each level has been put in there through the mill
of the conflict of evolution that has finely honed in an arbitrary
but nevertheless purposeful fashion.
SW: Now you're bringing in another word, that is a very slippery
word, and that is the word purpose. The question is what has a purpose
and how would we tell. For example one thought experiment that I
think is kind of interesting is again sort of an extraterrestrial
intelligence thought experiment which is we look at the configuration
of stars in the sky and we imagine a very advanced civilization
that's capable of moving stars around. The question is what
would they do to move the stars for a purposeful purpose, so to
speak? Would they make simple geometrical shapes? Turns out that
there are gravitational processes that will make perfect triangles
and so on. Would they make some more complicated kind of random
shapes, maybe the kind of shapes that the Babylonians identified
as constellations and we actually see.
RK: Yes, but triangles we see in the cellular automata and that
but I the one hand and it really gets to the heart how far CA takes
us. I was struck with the beauty in both order and randomness of
the output of these class 4 automata. On the other hand, all the
pictures in the book are recognizable as a certain level of what
I would call complexity. And in your doctrine of equivalence, all
these different levels of order, different levels of complexity,
are rendered equivalent.
SW: When you said all these different levels you are actually really
only identified two. You identified human intelligence and other
stuff. Is there yet other levels or are we talking about?
RK: I think there's a continuum; I mean you can take—
SW: What's an example of something, which is—
RK: In between? Say, current state of AI is in between. It can
perform functions that used to require human intelligence but it
doesn't yet have the suppleness and subtlety and range and
ability to deal with language of human beings. It's somewhere
in between. Many of our examples of technology are in between and
other animals, say, apes, you mentioned whales. We've identified
other species that seem to have some level of intelligence. There's
a sense that a whale, a squid maybe at our level, maybe below. Dogs
and cats we consider to be somewhat intelligent, more intelligent
than an insect, which may be more intelligent than a single cell
creature. So, we do have a sense of a continuum of intelligence.
SW: I think this question of what has a purpose and how can we
tell whether a thing has a purpose, when you're presented with
a thing how do you tell whether that thing was made for a purpose
or not? Well, one way that one can guess is if the thing achieves
some particular function, performs some particular function, there's
a question of whether it performs that function in some minimal
way or whether it performs that function in some elaborate sort
of ornamental kind of way and I think that this is a way of sort
of potential definition of purpose is to say that a thing visibly
has a purpose if it achieves it's function.
RK: [Inaudible] a machine's intelligence, the ability to use
in an optimal fashion limited resources including time to achieve
a set of goals.
SW: What I was going to say which now somehow disagrees with that
while in principle one can recognize a purpose by saying the thing
achieves what it achieves in a minimal way, our technology is very,
very far away from that. And, certainly our intelligence is very
far away from that. If we look at a Pentium chip or something.
RK: That's the point of evolution. It doesn't immediately
go to optimal. It's continually having different things compete,
and the ones that are more optimal, use less resources, achieve
the result a little bit better or faster. survive and the other
ones don't and that process continues so it continues to get
better and better and more and more optimal.
SW: But that's a common claim of kind of the evolutionary
doctrine. One of the things I've been interested in is the
question when we see complexity in biology, very obvious kind of
tangible complexity like elaborate pigmentation patterns on mollusk
shells, those sort of things, what can we say about what their origin
is. Is their origin some process of optimization? Some sort of careful
crafting by trying to fit into some ecological niche. Or is it just
various random programs were tried and many of those random programs
turned out to produce these complicated patterns. When I've
gone back and looked at some of those experiments the people have
done. They say, look we've got this very complicated stuff
going on. This is a sign that something very interesting is happening
in our evolutionary process. They're not right about that.
That this sort of, they're not seeing what they think they're
seeing which is the complexity they're getting is somehow carefully
crafted by this complicated process of evolution.
RK: There is a honing of details of a design through evolutionary
process.
SW: For sure.
RK: I think there's some problems with the way we set up our
evolutionary algorithms and we need ways of evolving the methods
of evolution. We didn't stay with one chromosome and things
evolve at every level, but I mean, how do you get from a class 4
automata, to the kind of elaborate hierarchy of design that you
see in a human being? They don't evolve just by themselves.
Just running the CA by itself you will continue to get the same
level of complexity.
SW: The universe doesn't evolve and yet within the universe
things like us come to be. If I'm right that there's a
simple program that represents the universe. If there is a definite
simple program that represents the universe. That program gives
everything. It gives this process.
RK: There's one other issue I wanted to get at on the level
of physics, the idea of particles presumably would be some sort
of glider in how we understand them in automata, and the glider
would embody the encoding of information to represent the properties
that we associate like spin and so on.
OP: Why is it called a glider?
RK: A glider because you have the cellular, the celestial computer
meaning the cellular automata, that is implementing these rules
or is the rules. And, so it comes from one state to the next and
the phenomenon that was discovered fairly early on Stephen's been
involved with CA's for a good twenty years is that you can have
certain aggregate pattern that actually copies itself into the next
state exactly but generally displaced let's say by one cell or maybe
diagonally by a cell and it will then move and so it glides across.
That kind of looks like a particle. If you have two gliders and
they hit each other they will interact kind of like fundamental
particles in an accelerator. It's suggested, anyway that you could
define fundamental particles as CA gliders and they embody in their
shape and so on a certain amount of information and Ed Fredkin has
said that a fundamental property is spin; it would be encoded somehow
in the definition of these gliders.
SW: This is kind of a naïve level of understanding of what
it means for the universe to be like a computer. This is kind of
a we've been talking a bunch about cellular automata and I
haven't made the more fuss each time to say no it isn't
just cellular automata it's simple programs in general, but
this is a place where that distinction matters. A cellular automaton
is a very specific kind of simple program which has as you say a
grid or an array of cells like the arrangements of atoms in a regular
crystal or some such other thing. The traditional idea that's
existed in most of physics that space just is and then there's
matter and all the particles and so on that do things on top of
space. Part of what I think is going on and it's more abstract
more difficult to understand and to explain is something where space
is all there is and it's features of space itself that correspond
to things like particles and so on. The analogy is something like
some fluid like water, seems to be continuous, just like space seems
to be to us continuous in the sense that you can move from anywhere
to anywhere in arbitrarily small increments and so on. But in fact
we know that water isn't at an underlying level a continuous
fluid, it has a bunch of discrete molecules bouncing around.
RK: There has to be some kind of network and our conception of
space is an abstraction where the fundamental reality is the cellular
network, and it I'm not saying it can't be done but its
not clear how you get a network that would give you the results
that we've seen.
SW: That part is actually very easy. There are other parts that
are hard. But that part is quite easy. The thing is that it's
easy but it's abstract so you have this network and it has
these nodes and it's connected to other nodes. At first you
might not have any reason to think that the sort of aggregate of
that when you have enough nodes that what you would get would be
anything that even vaguely approximates ordinary space. It's
a slightly long story that I talk about in the book. One has a notion
of space. One has a notion of time based on updating these networks
and the remarkable thing is that with certain kinds of underlying
rules, not only does one get invariance with respect to rotating
things around in space one also gets the more subtle invariants
that one gets, that' associated with special [inaudible] that
there's it's in a sense it's not surprising that
there's no directionally issue because this model of space
doesn't have anything like a regular grid. It has a network,
and it's an network that has random features and it's
not surprising that this network with many random features that
they won't be any particular preferred directions that exist.
I might say that defining what a particle is in the context of this
idea that there's nothing in this universe except space is
a tricky business. Actually one person who did think seriously about
idea that there might be nothing in the universe except space was
Einstein. In the later years of his life when he tried to develop
what he called the unified field theory which has little to do with
modern unified grand unified theories and so on he had the idea
that the only thing that might exist in the universe is gravity
and gravitational fields and that somehow the particles we see might
be some sort of knots or singularities in the gravitational fields.
He tried to make that work. He couldn't make that work on
the basis of continuum, Einstein equations and so on that he had.
I think what I believe is that in a minimal model for the universe
which I think is always the place to start from that is going to
work, that there might need to be nothing in the universe except
space and all the features, all the particles and so on that we
see might end up being just features of this space. The reason that
I think that the underlying stuff of the universe is based on these
networks is sort of the network has as little as possible built
into it; it doesn't have a notion of space. It doesn't
have a notion of colors of cells and it doesn't have a notion
of, it has sort of the minimal set of possible notions built into
it. It doesn't know how many dimensions it's in. It's
just a bunch of connectivity information and from that there then
can what's interesting is there can emerge from that notions
like space, like time, and to my surprise, special relativity and
to my even greater surprise, general relativity and features of
gravity.
OP: Stephen, do you see this having any practical applications?
SW: Sure.
OP: Like what.
SW: In fact, my more extreme, my sort of brash predication, okay
is that fifty years from now, of the new technology that is getting
produced at that time, more of it will be based, for example from
the kinds of ideas in my book, then is based on ideas that emerge
from calculus and traditional mathematical approaches to science.
OP: So researchers could actually take the work that's in
the book and use that to help them evolve new—
SW: Yes. Absolutely. I know there are a great many different kinds
of researchers and technology people who are trying to use things
in the book to a variety of different kinds—
OP: Like nanotechnology for instance?
SW: That's one kind of application. There are different kinds
of applications. Some applications are to essentially scientific
questions. Taking simple programs and using those as models for
specific kinds of things, whether in physics, or there's a
particular enthusiasm for doing that in biology. It's a clear
issue right now what is the appropriate way to idealize the processes
that go on in the biological cells, to make a theoretical computational
biology. I think there's considerable enthusiasm which I share
for using the kind of simple program mechanisms that I talk about
in the book as the raw materials to make appropriate models for
processes that go on in cells. And in other kinds of technology,
there are questions like if you want to make a computer out of the
minimal possible components, it's important to know just how
simple the rules can be to make the computer, and so for example
the systems that I look at that correspond to very simple rules
that turn out to be capable of universal computation. That's
important to know about because those things are things that you
might realistically be able to implement using atoms and molecules
and so on. That's another I think, important direction. At
first it might seem absurd to say we can implement everything using
one of these rules where you might have to build all this elaborate
software to use that rule, but if that rule can be implemented at
the level of atoms, a few layers of software doesn't really
cost you that much relative to what you gain having the thing implemented
at the level of atoms. Another quite different type of application
is for artistic purposes where often art is inspired by what we
actually see in nature and what we are familiar with in nature.
If one can understand the essence of what produces the forms that
we see in nature, that kind of expands the domain of what one can
use to do things for artistic or computer graphics and other kinds
of purposes like that.
RK: I think it's important work and in my mind it's significance
to be most important in what it tells us abstractly on a few key
points. One is the simplicity of computation and it's ubiquitousness.
When we get into nanotechnology where the entities are going to
be very simple. And we're going to try to build up complex
results from very simple devices to help them organize themselves;
we want very simple models of computation. However, I don't
think we can achieve, say the goal in my field, which is AI, strong
AI, which is the ultimate goal, roughly defined as replicating the
full range of human intelligence, understanding and reverse engineering
recreating it's basic principles, just through CA's. Starting
with simple conditions and jumping to the complexity you get with
rule 130. It's going to jump, but not to strong AI. We're
going to need to understand what I've been calling the cascade,
the hierarchy of features that was the arbitrary result of billions
of years of evolution that produced human intelligence.
OP: However we get there, though, both of you earlier agreed that
you think machines eventually will, at least—
RK: And not that long from now in my mind. Within three decades.
SW: I would say that the idea that from a rule 110 evolving on
a computer, suddenly an intelligent, human like intelligence will
jump out, this will not happen. We all know this will not happen.
The question is we talked about the much more abstract issue of
whether there's an essence of intelligence that is not captured
by processes like the ones going on with rule 110, like the way
that rule 110 would sort of respond to it's environment. Is
there something really qualitatively and essentially different about
that then about the way that humans do it? That's a separate
issue from whether some of the kinds of ideas about simple programs
and what they do can be useful in the practical pursuit of AI. You
said that you think that there is some essence of intelligence that
is not specific to our brain that is some kind of abstract essence
of intelligence yet you seem to think that to achieve AI, the only
way we're going to be able to do that is by reverse engineering
what the actual brain that sort of historically—
RK: Or any brain of that level of complexity that we can get our
hands on.
SW: It's still a brain.
RK: That's the only one we know.
SW: But you see I think that in a sense—
RK: If we could find another one we could reverse engineer that.
SW: In a sense that kind of betrays what's true about the
nature of intelligence. What you're basically saying is the
way that we do this thing that sort of captures intelligence is
to take this one example that we think we definitely have which
is human brains and reverse engineer that. We don't sort of
abstractly think about sort of hierarchies of whatever and use that
to kind of build up our intelligent thing.
RK: We do do that as well. Most of the history of AI has not been
reverse engineering of the human brain. Most of the history to date
has in fact been creating a different kind of intelligence that's
also has a certain amount of hierarchy but is a result of our engineering
just as our flying machines aren't emulations of birds. And,
both of those methods are going to continue to get us make progress.
OP: Two years ago of course, Bill Joy took one look at this situation
and threw up his hands and said we're in danger of killing
off the human race because we could create entities nanorobots that
would—
RK: The debate between Bill Joy and myself started from Bill Joy's
reaction to a conversation we had in September of 1998 at a George
Gilda conference and I gave him an advanced copy of The Age of
Spiritual Machines, which came out a few months later. Where
we do disagree is on the feasibility and desirability of relinquishment.
Nanotechnolgy is not being done in a few nanotechnology labs is
simply the inevitable end results of the miniaturization of technology
that exists across the board. When Texas Instruments makes a higher
resolution projector or a better camera, that's a step towards
nanotechnology. Each one of them is benign, it's thousands
of benign steps that gets us to these future technologies and we're
going to have to put specific effort into the defensive technologies,
just as we do with computer viruses. In fact if we do half as well
with viruses software viruses as we in these other areas as say
biological viruses or nanotechnology entities we'll do well.
People say let's get rid of the Internet because of software
viruses, but I think it is a major challenge facing humanity. Look
at the destruction in the 20th century. Two world wars killed almost
a hundred million people made possible by technology. 21st century
could be worse. It also has the potential to alleviate age-old suffering.
Our technology will have the means to overcome disease, extending
longevity, overcoming poverty, and cleaning up the environment.
OP: Do you see Stephen's work contributing more to this pro
or con?
RK: I think it's a powerful paradigm that simple methods can
give more complex results. You can have lots of simple methods interacting
with each other and through the sort of collective interaction get
interesting results.
SW: I'm looking forward to that.
OP: Me too, frankly. I don't think I'm going to be here
long enough.
RK: It's an important work and in insights it will help us
along the way.
OP: Are you concerned?
SW: I'm an optimist and I also think that the idea of saying
lets make some rule, some law that we say you can't do this
big swath of poorly designed things is unlikely to have good results.
RK: It will just drive it underground.
OP: If you're an optimist what saves humanity? Once we have
intelligent machines that can replicate themselves in biological,
our method of reproduction will seem like a snail's pace.
SW: It's an interesting question. What will the end point
of the human condition will end up being. There will be a time,
I think and Ray agrees about this that there could be a time when
we could have this little cube sitting on the desk that would replicate
the thought processes of any of us.
RK: Or all of us.
SW: And the question is for example, what would these cubes sitting
around, what would they do? What will be, many of the things that
have shaped human civilization are various kinds of constraints
that we've tried to push beyond and these sorts of things,
what would these little cubes just sitting around having all the
resources they want, what will be what they choose to do? I don't
think we know, necessarily, I think it's also the case that
in the end it's sort of a disappointing conclusion that we
can take all of our humanity and wrap it up in these little solid
state cubes.
RK: When you describe it that way it sounds like it's minimizing
it and whenever you talk about people becoming machines, people
think about the machines they've known and they don't
want to be a machine like that because that's going to be a
step backwards because these machine are very primitive. In my mind
it's not an alien invasion of intelligent machines coming from
over the horizon to replace us but it is really emerging from within
our civilization. We'll start out by merging with it quite
literally. We're now pretty close to our technology. We keep
it in our pocket and ultimately some of us are already putting it
in our brains and bodies, generally people with disabilities, medical
problems, but there are a whole bunch of neural implants that will
replace brain circuits and will work with biological neurons and
ultimately we will expand what it means to be human and ultimately
the non-biology, non biological portion will dominate because it's
growing exponentially whereas the biological capacity is relatively
fixed.
SW: This is a place where we agree.
RK: But they've got to be doing human things and they've
got to be doing human things more than we do it. I'd like to
be more human. I'd like to be able to express myself more eloquently
at times, and I'd like to be able to create music at a certain
level. I'd like to be able to create more beautiful music.
I'd like to be able to do things at the peak of human ability
all the time and so on and these technologies will enable us to
be more of the positive qualities that we associate with humanity.
Our conflicts will also be more intense but it will amplify what
it means to be human.
OP: Getting much less grandiose, you even used you even explored
the potentials for encryption, what about decryption?
SW: You mean cryptanalysis for example?
OP: Yes.
SW: I never thought about this question how would what I have done
apply to cryptanalysis, but since you asked the question, the I
think that the idea of being able to enumerate all possible simple
programs and see what they do has some potential in cryptanalysis.
OP: The work that you've done has this enormous breadth of
application from very mundane things like encryption and decryption
all the way up to revising our understanding of the structure of
the whole cosmos. If all this proves out is this going to be the
most important book of the decade?
SW: The ideas are important. The notion that simple programs and
what they do this is an important thing. This book happens to be
the encapsulation of this idea. The idea is the thing that is important
and will be a kind of in the history of science for example, will
end up being seen as one of the more important ideas.
OP: Stephen very interesting. I thank you very much for coming.
It was enjoyable.
SW: Thank you.
OP: And for helping me understand a little more about what Stephens'
been doing. Thanks very much, Ray.
RK: My pleasure.
© 2005 Otis Port/Wolfram/Kurzweil. Reprinted with permission.
| 
