Computers, Science, and Extraterrestrials: An Interview with Stephen Wolfram
Extract from: David G. Stork, ed., HAL's Legacy: 2001's Computer as Dream and Reality, MIT Press, 1996.
DGS: If I calculated correctly, 2001 must have come out when you were nine years old. Did you see it then?
SW: Definitely. In fact it was my favorite movie, and I ended up seeing it quite a few times.
DGS: What was it that you liked about it?
SW: All the technology. It had such a rich rendering of technology. So bright and attractive. I followed the U.S. space program quite a bit in those days, but I was always disappointed when I realized that the inside of a spacecraft was just the size of a closet. In 2001 there were masses of technology, and I really liked it.
DGS: Did you identify with any of the characters in 2001?
SW: Not really. At the time, I was young, and they all seemed quite old, and very American. Growing up in England, all that "Roger your plan to go E V A" stuff seemed pretty foreign to me. I guess I didn't ever imagine talking jargon like that. One thing though was that I definitely was very interested in one day being able to interact with the kind of technology that was in the movie. At that time, though, I thought that meant that I'd have to join NASA or something. I'm sure I didn't think about it in those terms, but I certainly didn't imagine the kind of flattening of who has access to technology that's actually happened.
DGS: I gather you just watched 2001 again this evening. What struck you most about the movie now?
SW: Well, it seemed a lot shorter than when I was a kid. It also seemed to make a lot more sense. I guess I now at least think I understand how all the pieces are supposed to fit together. And they really raise some very interesting questions. But let's talk about that later.
Another thing that struck me a lot watching 2001 this evening was how emotionally cold it is. I don't think I noticed that at all when I was a kid. But seeing the movie now I was just amazed at how little human emotion there is in it. I think HAL was the most emotive of the bunch in the end.
When I was a kid I guess I had this vague idea that scientists were the kind of emotionless characters that are portrayed in the movie. They never seemed to get excited about anything. Maybe that's a bit of what's wrong with science, actually. For most people there's so little passion in it. Just get the data, analyze it from 9 to 5, and so on. Kubrick may have had that exactly right.
Anyway, you ask what struck me about the movie. Another thing is how rich almost every scene is. There were so many details. So many things that someone must have really thought about. It's a damned good piece of work. And I suppose what's nice about it is that enough people think that every scene has really been studied carefully. And people are even writing books about what happened in the movie. Like this one, I guess!
DGS: A lot of things haven't worked out exactly as 2001 predicted. Does that surprise you?
SW: Well, given the level of detail in the movie, it's an absolute set-up to be proved wrong. I've got to say that I'm really impressed by how much was got right. And I think a lot of the mistakes are really interesting mistakes—mistakes that one learns something by seeing why they were made. I actually think quite a lot about trying to predict things, and I find it incredibly useful to go back and see why mistakes in predicting got made before.
I suppose I've noticed two big reasons for mistakes. One is historical accidents—things that worked out the way they did for some fairly chance reason: because some particular person invented something in a particular way or whatever. And another is much deeper: not understanding a basic concept. Not getting a big idea—a paradigm shift or whatever—that really changes how a lot of stuff works.
Of course in looking at the movie some of the most obvious mistakes aren't about technology. The hairstyles look all wrong. The voiceprint identification system asks for a "Christian name" and so on. Actually it's interesting that the various companies portrayed in the movie—Bell, Pan Am, BBC, Hilton, and so on—are all still around, at least in one form or another. But most of them have quite different logos now. The fact that graphic design tastes change is a pretty general thing; but which particular companies changed their logos is definitely in the category of historical accident. Same with the fact that some of the typefaces in the movie look dated, and some don't.
DGS: What about the computers in 2001?
SW: Well, let's talk about the ordinary ones—not HAL—for now. It's really fascinating what was predicted correctly there, and what wasn't. There was one definite major conceptual mistake, I think, that had to do with misassessing the power of software, and that pervaded a lot of the things that weren't got right.
One thing that was got very right is that computers would be important, and that there would be computers—or at least computer screens—everywhere. But the thing that was got wrong was how much stuff would have to be done with different special-purpose devices, and how much could be done just with software. There were fine in-flight TV screens on the shuttle. But they had rows of separate buttons underneath, not software menus. There were lots of separate computer screens showing different things. Not just software-controlled windows. People were looking at physical photographs of the monolith on the moon, not computer renderings. And there was a click for the camera shutter. Not digital. And of course there were clipboards being used on Discovery.
Now, one can argue that airplane cockpits still have rows of buttons, and that the clipboard thing was just about not predicting portable computers. But I think there was more to it. I think Kubrick and Clarke didn't have the idea that once you're dealing with a general-purpose computer, you can do things purely in software, without having to have different special-purpose hardware.
I certainly don't blame Kubrick and Clarke for making the mistake. In fact, people often still make the same mistake today. But if one looks at the history of computing there's an extremely clear trend: special-purpose hardware gets replaced by software running on general-purpose machines. One doesn't need physical teletypes anymore, because the forms of letters can be made in software. Soon one won't need video hardware, because all the signal processing will be able to be done in software. People often don't see it, but universal computers really are universal. And it's only a matter of time before pure software can do more and more things. Without needing special hardware stuff.
It's funny. Watching 2001 really makes me think about the significance of universal computing and software. 2001 in a sense makes the case that it was the invention of tools that really got humans started on the path to where they are today. Well, I guess in the last few years I've come to think that the invention of software is something of about the same magnitude as the invention of tools. You see, before you have tools, the only device for getting things done is your own body. But with tools, you can go beyond that. Still, once you've built a tool, you're stuck with that particular tool. The idea of a universal computer is that you can make a universal tool—a general-purpose object—that you can program to do absolutely anything. And I think now that we've only just started down the path that's opened up by the idea of software. There's probably as much development to come as in the sequence in 2001 that cuts from a bone as bashing tool to an orbiting spacecraft.
DGS: Let's talk for a bit about scientific computing. It seems like that's an area where reality has surpassed prediction.
SW: Probably so. Mathematica can certainly make much nicer displays than the ones in 2001. Particularly with Mathematica 3.0, which, with some luck, should be out long before HAL's 1997 birthday. And there's probably even been at least one Mathematica in space by now—on a portable computer on the shuttle, or something.
But I guess the clearest piece of progress that's been made relative to what one sees in 2001 is in the language that's getting used for scientific computing. If one looks at the closeups of screens in 2001, one sees code that looks like BASIC or Fortran. It's very different-looking from Mathematica code—particularly from Mathematica 3.0 code.
I'm sure nobody thought very hard about it when the screens for 2001 were made. They probably just took something that looked like computer programs that existed in the mid-1960s. But in fact there's been a very important change. You see, in the mid-1960s computer time was at a premium, so you had to make things as easy as possible for the computer. The result is that the languages were optimized for the way the hardware of the computer would process a problem—not for the way people might think about a problem.
It's kind of like the underestimation of the power of software that I mentioned before. People assumed that you'd have to have computer languages that fitted into the structure of the hardware that existed. They didn't understand that you could make layers of software that would make the languages work more like the way people think—or the way that, for example, math works.
And when I built Mathematica, one of my big ideas was to try to make a language that wasn't just tied to hardware—and instead that was set up to work the way people actually think about computations.
Now, here's the bad part: I think Mathematica was probably in a sense just an accident. You see, computer languages seem to live an incredibly long time. And all those languages like Fortran and BASIC that were developed in the 1950s and 1960s are still alive today. They've long outlived all the computers they were originally developed for, and all the issues they originally had to address. But they're still around. And I'm guessing that if Mathematica hadn't come along, they'd still be the dominant languages for scientific computing—even in 2001. So I suppose I'm guessing that the differences between the scientific computing in 2001 and in reality are probably pretty much a historical accident—albeit in a sense my personal historical accident.
DGS: What about issues of basic science? Are there things in 2001 that you found interesting from that point of view?
SW: Actually, very much so. I thought the portrayal of extraterrestrial intelligence was really fascinating. It's all linked up with questions about computation in nature and the distinctions between natural and artificial things. In fact, I've got some stuff on extraterrestrial intelligence in the book called A New Kind of Science that I've been working on for the past few years. But 2001 has some very nice examples of the issues.
I guess the big question is how one knows when something is "natural'' and when it's "artificial''. How do we know when a signal we get from the cosmos is just "natural'', and when it's "artificial'', and made by some kind of intentional being?
At the beginning of 2001, we see all sorts of obviously natural stuff—mountains, skeletons, apes running around. But then suddenly we see the monolith—and it's obviously artificial. How can we tell? Well, unlike all the other stuff we've seen, it has perfectly smooth sides, and looks like something that's been engineered, and hasn't just grown naturally. Actually, that's the heuristic we essentially always use: if something looks simple, then it's probably artificial, and if it looks complicated, then it's probably natural.
I happen to think that that heuristic is a very clear sign of one of the biggest shortcomings of present-day science and technology. You see, what it's saying is that there is some kind of secret ingredient that nature is adding to make stuff complex—and that we don't know from a scientific point of view. Well, actually, for the past fifteen years I've been working hard to try to find that ingredient, and in a sense my new book is exactly about what the ingredient is.
The whole thing is a long story—that's why it takes a whole book to explain it—but the essence of it is that it's our reliance on mathematical equations and traditional mathematics that have made us miss what's going on. If instead of using mathematical equations, one thinks about things in terms of simple computer programs, then one can quite quickly see what's going on. And I guess that the big discovery that I made in the early 1980s is that there are some very simple computer programs that can do very complex things—things that are just like the things we see in nature.
Now here's the good bit: it turns out that those simple computer programs can often behave like universal computers. And what that means is that they can do stuff that's as complicated as anything, including anything our fancy electronic computers do. There's a major new piece of intuition here. You see people have tended to assume that to make a universal computer—a general-purpose machine—the thing had to be constructed in a pretty complicated way. Nobody expected to find a naturally occurring universal computer lying around. Well, that's the thing I've found isn't true. There are very simple universal computers. In fact, I think that lots of systems we find all over the place in nature can act as universal computers.
What does this mean? We can think of the behavior of any system in nature as being like a computation: the system starts off in some state—that's the input—then does its thing for a while, then ends up in some final state—which corresponds to the output. When a fluid flows around an obstacle, let's say, it's effectively doing a computation about what its flow pattern should be. Well, how complicated is that computation? It certainly takes quite a lot of effort for us to reproduce the behavior by doing standard scientific computing kinds of things. But the big point I've discovered is that this isn't surprising: the natural system itself is in effect doing universal computation, which can be as complicated as anything.
So in other words all our big fancy computers—and our brains for that matter—really aren't capable of computations that are any more sophisticated than a piece of fluid. It's a humbling conclusion—sort of the ultimate put-down—after we discover the earth isn't the center of the universe, that our bodies work mechanically, and so on. But it's very important in thinking about extraterrestrial intelligence.
Here's why: if we receive a signal, we've got to figure out whether it came from something intelligent or not. Now here's where the issue about natural versus artificial gets very confusing. If the thing we see is too simple, we're probably going to conclude it's not coming from anything intelligent. Actually, if we'd only seen the monolith—and it hadn't been quite as big as it was—we might have concluded that it was just a crystal—some very fine specimen of a black gemstone.
Likewise, when Dave enters the stargate, we first see just a plain row of lights, which could easily come from some simple physical process. Later, things start looking more complicated. And then, at times, things look sufficiently random and complicated that we would probably conclude that they were "just natural'', and not artificial or intelligently created in any way.
So actually the stargate sequence is a very good example of how difficult it can be to tell whether something is natural or artificial—whether it has been made "intentionally'', or just grown naturally. I'm not even sure what Kubrick had in mind in parts of that sequence. Later on in the sequence, we seem to be over a natural planet surface, but Kubrick added some flashing octahedra just to make it clear that the whole thing wasn't just supposed to be completely natural—unless perhaps those octahedra were crystals, or something.
DGS: Actually, the octahedra were Kubrick and Clarke's extraterrestrials—sort of escorts bringing Dave through the stargate. We can be very thankful indeed that they threw out the version of the screenplay that had a New York ticker-tape parade with the octahedra riding along in convertible cars!
SW: Gosh. And there I was thinking that the octahedra were just supposed to be simple beacons, flashing like lighthouses or something. Not intelligent objects at all. I guess that just shows how difficult it can be to tell whether something is supposed to be intelligent or not!
Anyway, all this stuff about the natural versus the artificial definitely isn't just of theoretical interest. After all, we have picked up all sorts of mysterious radio signals from the cosmos. There were some very regular ones first discovered within just a few weeks of when 2001 came out: every few milliseconds an intense radio burst arriving from the Crab nebula, and other places. And I guess at first people thought that perhaps this was a sign of extraterrestrial intelligence—but then they realized it was just a communal garden variety natural neutron star. Well, now we know that the radio pulses from neutron stars aren't actually perfectly regular: they have little modulations and so on. Is this a sign of extraterrestrial intelligence? A signal superimposed on a carrier? It's incredibly unlikely—a much better theory is that the modulations come from quakes in the crust of the neutron star. And one reason one might conclude that is that the modulations seem fairly random: they seem too complicated to be artificial. But then, of course, if they were simpler, we'd probably assume they weren't produced by much of an intelligence.
DGS: So what kind of thing would make us sure we had detected extraterrestrial intelligence? What about receiving the digits of pi?
SW: Well, that's a tough one, for two reasons. First, how would we know that there was a complicated intentional intelligence generating those digits? You see, I've found some very simple systems that generate things like the digits of pi. Systems so simple that we could easily imagine they'd occur naturally, without intentional intelligence. So even if we found the digits of pi, we'd have a hard time being sure that the thing that produced them was something like us—a really complicated, evolved, learning thing—rather than just something like a piece of fluid.
Then there's a whole other problem: how would we know that we were receiving the digits of pi? You see, the digits of pi seem effectively random for essentially all purposes. And we certainly don't have any idea how to build some kind of analyzer that would systematically detect the digits as non-random. Well, the obvious question is: are there in fact radio signals that could be the digits of pi coming from around the galaxy. The answer is definitely yes. If you point a radio telescope in almost any direction, you'll hear "radio noise". Maybe it's all thermal emission from hot gas, but maybe—just maybe—there are the digits of pi out there. We don't right now have any way to know for sure.
DGS: So, with your definitions, do you think there is extraterrestrial intelligence out there?
SW: Oh, I'm sure there are lots and lots of systems that can do computations as sophisticated as working out the digits of pi. We've got lots right here on earth. But we don't call them intelligent. Even though some of them seem to "have a mind of their own"—like the weather. But I also think there are probably lots of extraterrestrials out there of the kind you're talking about—with lots of history, evolution stuff, and so on.
DGS: Will we find them?
SW: I expect so. And probably eventually the argument about whether the signals we get from them are really "natural" or "artificial" will die down. But my guess is that history will work out so that we build artificial intelligence in computers before we find extraterrestrial intelligence. And the result of that is that finding extraterrestrial intelligence will be considerably less dramatic to us. Because by then we'll already know that we're not the only intelligent things that can exist.
DGS: So what about HAL? It's almost 1997 and we don't have anything like HAL. Why do you think that's happened?
SW: Probably you'd expect me to say it's because our computers aren't fast enough, thinking is a difficult thing to get, and so on. But I really don't think so. I think it's just a historical accident.
Sometime—perhaps ten years from now, perhaps twenty-five—we'll have machines that think. And then we'll look back on the 1990s and ask why the machines didn't get built then. And I'm essentially sure the reason won't be because the hardware was too slow, or the memories weren't large enough. It'll just be because the key idea or ideas hadn't been had. You see, I'm convinced that after it's understood, it really won't be difficult to make artificial intelligence. It's just that people have been studying absolutely the wrong things in trying to get it.
The history of AI is quite interesting: I think in many ways it's a microcosm of what's wrong with science and academia in general. Everyone knows that when computers were first coming out in the 1940s and 1950s many people assumed that it'd be quite easy to make artificial intelligence.
The early ideas about how to do it were, in my view, pretty sensible, at least as things to try—simple neural nets, stuff like that. But they didn't work very well. Why not? Probably mostly because the computers in those days had absolutely tiny memories. And to do anything that remotely resembles what we call thinking one has to have a fair amount of knowledge—and that takes memory. Of course, now any serious computer can easily store an encyclopedia—so now that problem should have gone away.
Well, anyway, after the failures of the early brute-force approaches to mimicking brains and so on, AI entered a crazy kind of cognitive engineering phase—where people tried to build systems which mimicked particular elaborate features of thinking. And basically that's the approach that's still being used today. Nobody's trying more fundamental stuff. Everyone assumes it's just too difficult. Well, I don't think there's really any evidence of that. It's just that nobody has tried to do it. And it would be considered much too looney to get funded or anything like that.
DGS: So what kind of approach do you think will work in building intelligent machines?
SW: I don't know for sure. But I'm guessing that a key ingredient is going to be seeing how computations emerge from the action of very simple programs—the kind of thing that happens in the cellular automata and other systems I've studied. I think that trying to do engineering to mimic the high-level aspects of thinking that cognitive scientists or psychologists have identified is not going to go anywhere. Thinking is, I'm pretty sure, a much lower-level process. All those cognitive things are just icing on the cake—not fundamental at all. It's like in a fluid: there are vortices that one sees. But these vortices are not fundamental. They are a complicated consequence of the microscopic motions of zillions of little molecules. And the point is that the rules for how the vortices work are fairly complicated—and hard to find for sure. But the rules for the molecules—they're fairly simple. And I'm guessing that it's the same way with the underlying processes of thinking.
DGS: So do you really think that we can get a handle on profoundly hard, high-level problems of AI—such as my favorite, scene analysis—by looking at something as "simple" as cellular automata?
SW: Definitely. But it takes quite a shift in intuition to see how. In a sense it's about whether one's dealing with engineering problems or with science problems. You see, in engineering, we're always used to setting things up so we can explicitly foresee how everything will work. And that's a very limiting thing. In a sense you only ever get out what you put in. But nature doesn't work that way. After all, we know that the underlying laws of physics are quite simple. But just by following these laws nature manages to make all the complicated things we see.
It's very much connected with the things I talk about in A New Kind of Science. It took me more than ten years to understand it, but the key point is that even though their underlying rules are really simple, systems like cellular automata can end up doing all sorts of complicated things—things completely beyond what one can foresee by looking at their rules, and things that often turn out to be very much like what we see in nature.
The big mistake that gets made over and over again is to assume that to do complicated things one has to set up systems with complicated rules. That's how things work in present-day engineering, but it's not how things work in nature—or in the systems like cellular automata that I've studied.
It's kind of funny: one never seems to imagine how limited one's imagination is. One always seems to assume that what one can't foresee can't be possible. But I guess that's where spending fifteen years doing computer experiments on systems like cellular automata instills some humility: over and over again I've found these systems doing things that I was sure wouldn't be possible—because I couldn't imagine how they'd do them.
It's like bugs in programs. One thinks a program will work a particular way, and one can't imagine that there'll be a bug that makes it work differently. I guess intuition about bugs is a pretty recent thing: in 2001 there's a scene where HAL talks about the fact that there's never been a "computer error" in the 9000 Series. But the notion of unforeseen behavior that isn't due to hardware malfunction isn't there.
Anyway, about "hard" problems in AI: my own very strong guess is that these will be solved not by direct engineering-style attacks, but rather by building things up from simple systems that work a bit like cellular automata. It's somewhat like the hardware versus software issue we discussed earlier: in the end I don't think elaborate special-purpose stuff will be needed for problems like scene recognition; I think they'll be fairly straightforward applications of general-purpose mechanisms. Of course, nobody will believe this until it's actually been done.
DGS: So, have you yourself worked much on the problem of building intelligent machines?
SW: Well, since you ask, I'll tell you the answer is yes. I don't think I've ever mentioned it in public before. But since you asked the right question: yes, I have been interested in the problem for a very long time—probably twenty years now—and have been steadily picking away at it. I've been held back by a lack of tools: both practical and conceptual. But that's finally getting sorted out. I have Mathematica from the practical side to let me do experiments easily. And I have my new science, from which I think I've figured out some of the basic intuition that's needed. And I even have my company—headquartered in Champaign-Urbana, birthplace of HAL, as chance would have it—that can potentially support my efforts. But I guess I'll have to disappoint you: we won't be announcing a machine that thinks in 1997. It'll just be Mathematica version X and A New Kind of Science from me. But wait for another year, though. Perhaps 2001...