Originally published December 4, 2001 at Edge. Published on KurzweilAI.net December 4, 2001.

One of the striking things about being a computer scientist in this age is that all sorts of other people are happy to tell us that what we do is the central metaphor of everything, which is very ego gratifying. We hear from various quarters that our work can serve as the best understanding – if not in the present but any minute now because of Moore’s law – of everything from biology to the economy to aesthetics, child-rearing, sex, you name it. I have found myself being critical of what I view as this overuse as the computational metaphor. My initial motivation was because I thought there was naive and poorly constructed philosophy at work. It’s as if these people had never read philosophy at all and there was no sense of epistemological or other problems.

Then I became concerned for a different reason which was pragmatic and immediate: I became convinced that the overuse of the computational metaphor was actually harming the quality of the present-day design of computer systems. One example of that, the belief that people and computers are similar, the artificial intelligence mindset, has a tendency to create systems that are naively and overly automated. An example of that is the Microsoft word processor that attempts to retype what you’ve just typed, the notion of trying to make computers into people because somehow that agenda of making them into people is so important that if you jump the gun it has to be for the greater good, even if it makes the current software stupid.

There’s a third reason to be suspicious of the overuse of computer metaphors, and that is that it leads us by reflection to have an overly simplistic view of computers. The particular simplification of computers I’m concerned with is imagining that Moore’s Law applies to software as well as hardware. More specifically, that Moore’s Law applies to things that have to have complicated interfaces with their surroundings as opposed to things that have simple interfaces with their surroundings, which I think is the better distinction.

Moore’s Law is truly an overwhelming phenomenon; it represents the greatest triumph of technology ever, the fact that we could keep on this track that was predicted for all these many years and that we have machines that are a million times better than they were at the dawn of our work, which was just a half century ago. And yet during that same period of time our software has really not kept pace. In fact not only could you argue that software has not improved at the same rate as hardware, you could even argue that it’s often been in retrograde. It seems to me that our software architectures have not even been able to maintain their initial functionality as they’ve scaled with hardware, so that in effect we’ve had worse and worse software. Most people who use personal computers can experience that effect directly, and it’s true in most situations.

But I want to emphasize that the real distinction that I see is between systems with simple interfaces to their surroundings and systems with complex interfaces. If you want to have a fancy user interface and you run a bigger thing it just gets awful. Windows doesn’t scale.

One question to ask is, why does software suck so badly? There are a number of answers to that. The first thing I would say is that I have absolutely no doubt that David Gelernter’s framework of streams is fundamentally and overwhelmingly superior to the basis in which our current software is designed. The next question is, is that enough to cause it to come about? It really becomes a competition between good taste and good judgment on the one hand, and legacy and corruption on the other – which are effectively two words for the same thing, in effect. What happens with software systems is that the legacy effects end up being the overwhelming determinants of what can happen next as the systems scale.

For instance, there is the idea of the computer file, which was debated up until the early 80s. There was an active contingent that thought that the idea of the file wasn’t a good thing and we should instead have a massive distributed data base with a micro-structure of some sort. The first (unreleased) version of the Macintosh did not have files. But Unix jumped the fence from the academic to the business world and it had files, and Macintosh ultimately came out with files, and the Microsoft world had files, and basically everything has files. At this point, when we teach undergraduates computer science, we do not talk about the file as an invention, but speak of it as if it were a photon, because it in effect is more likely to still be around in 50 years than the photon.

I can imagine physicists coming up with some reasons not to believe in photons any more, but I cannot imagine any way that we can tell you not to believe in files. We are stuck with the damn things. That legacy effect is truly astonishing, the sort of non-linearity of the costs of undoing decisions that have been made. The remarkable degree to which the arrow of time is amplified in software development in its brutalness is extraordinary, and perhaps one of the things that really distinguishes software from other phenomena.

Back to the physics for a second. One of the most remarkable and startling insights in 20th century thought was Claude Shannon’s connection of information and thermodynamics. Somehow for all of these years working with computers I’ve been looking at these things and I’ve been thinking, “Are these bits the same bits Shannon was talking about, or is there something different?” I still don’t know the answer, but I’d like to share my recent thoughts because I think this all ties together. If you wish to treat the world as being computational and if you wish to say that the pair of sunglasses I am wearing is a computer that has sunglass input and output- if you wish to think of things that way, you would have to say that not all of the bits that are potentially measurable are in practice having an effect. Most of them are lost in statistical effects, and the situation has to be rather special for a particular bit to matter.

In fact, bits really do matter. If somebody says “I do” in the right context that means a lot, whereas a similar number of bits of information coming in another context might mean much less. Various measurable bits in the universe have vastly different potentials to have a causal impact. If you could possibly delineate all the bits you would probably see some dramatic power law where there would be a small number of bits that had tremendously greater potential for having an effect, and a vast number that had very small potentials. It’s those bits that have the potential for great effect that are probably the ones that computer scientists are concerned with, and probably Shannon doesn’t differentiate between those bits as far as he went.

Then the question is how do we distinguish between the bits; what differentiates one from the other, how can we talk about them? One speculation is that legacy effects have something to do with it. If you have a system with a vast configuration space, as is our world, and you have some process, perhaps an evolutionary process, that’s searching through possible configurations, rather than just a meandering random walk, perhaps what we see in nature is a series of stair steps where legacies are created that prohibit large numbers of configurations from every being searched again, and that there’s a series of refinements.

Once DNA has won out, variants of DNA are very unlikely to appear. Once Windows has appeared, it’s stuck around, and so forth. Perhaps what happens is that the legacy effect, which is because of the non-linearity of the tremendous expense of reversing certain kinds of systems. Legacies that are created are like lenses that amplify certain bits to be more important. This suggests that legacies are similar to semantics on some fundamental level. And it suggests that the legacy effect might have something to do with the syntax/semantics distinction, to the degree that might be meaningful. And it’s the first glimmer of a definition of semantics I’ve ever had, because I’ve always thought the word didn’t mean a damn thing except “what we don’t understand”. But I’m beginning to think what it might be is the legacies that we’re stuck with.

To tie the circle back to the “Rebooting Civilization” question, what I’m hoping might happen is as we start to gain a better understanding of how enormously difficult, slow, expensive, tedious and rare an event it is to program a very large computer well; as soon as we have a sense and appreciation of that, I think we can overcome the sort of intoxication that overcomes us when we think about Moore’s Law, and start to apply computation metaphors more soberly to both natural science and to metaphorical purposes for society and so forth. A well-appreciated computer that included the difficulty of making large software well could serve as a far more beneficial metaphor than the cartoon computer, which is based only on Moore’s Law; all you have to do is make it fast and everything will suddenly work, and the computers-will-become-smarter than-us-if-you just-wait-for-20-years sort of metaphor that has been prevalent lately.

Continued at Edge.

Copyright © 2001 by Edge Foundation, Inc.

www.edge.org