by K. Eric Drexler
From L5 News, November-December 1984
As space advocates, we all know that space technology does not stand alone: it is rooted in technology as a whole. Spacecraft did not give us advanced materials, computers, and computer-aided design — they gave us spacecraft. Although the challenge of space has accelerated development of these more basic technologies, they were a natural consequence of humanity’s broader effort to push hack the frontiers of the possible.
This broader effort, driven by global military and economic competition, continues at a quickening pace. To understand the prospects for space development, we must consider the broader prospects for technology. Two areas in particular seem crucial: automated engineering and molecular technology.
We have already seen great strides toward automated engineering under the name of computer-aided design (CAD). CAD systems have grown out of earlier structural-analysis programs and computer-aided drafting systems. Modern CAD systems let engineers create and edit designs on a computer screen. The computer stores a three-dimensional description of the design, and can provide the engineer with a picture of it from any angle. Many systems also allow the engineer to ask for an analysis of the object, not merely stating its mass, center of gravity, and so forth, but computing how it will bend under load and respond to heating. Computer models can often he tested in place of physical models, at a great savings in time and cost — and as computer costs fall and capabilities rise, this will become ever more common.
What is more, many CAD systems can take a final design and generate instructions that will direct computer-controlled machine tools to make the part, saving time and cost when the time comes to move from computer model to hardware. As automation and robotics increase, this, too, will become ever more common.
These systems automate part of the engineering task, but they do not yet amount to “automated engineering.” Engineers still must generate the ideas, however much work the machine may do in testing them. Thus, the machine cannot be said to evolve designs on its own. Other systems point the way to greater automation.
As usual, programmers have developed the best software tools for use in their own industry — computers. So-called silicon compilers can now take a specification of a proposed integrated circuit’s function and design a chip that will perform that function. They manage this in part because integrated circuits have only a few different device-types, combined according to a limited set of rules. Nonetheless, future automated engineering systems will extend similar abilities to ever-broader and more complex domains of engineering.
So-called expert systems show how a slice of a human expert’s knowledge can he described and embodied in software. A program for diagnosis in internal medicine (INTERNIST) has shown performance comparable to that of trained physicians, though only within a narrow field of competence. Another expert system (MYCIN) successfully diagnoses blood infections and prescribes treatments. Eventually, ever-larger slices of expert engineering knowledge will he similarly captured, for use by other engineers in diagnosing design problems and prescribing solutions. Programs resembling expert systems will increasingly suggest general design approaches. Programs resembling silicon compilers will increasingly assume the burden of filling design details. Upgraded versions of CAD modeling systems will increasingly simulate the resulting designs, to allow testing and revision without cutting metal.
Neither CAD systems nor silicon compilers nor expert systems now learn from experience. That is, while they can remember and use specific facts in the task at hand, they cannot draw new, general rules for use in future tasks. EURISKO, by Professor Douglas Lenat of Stanford, is different. It uses heuristics (rules of thumb, rules for plausible guessing) to select tasks, to suggest approaches to try, and to judge the results of the effort. It uses heuristics to suggest which heuristics to try, to judge which heuristics are working well, to warn when new heuristics are needed, and to suggest how to modify old heuristics to generate new ones. In short, EURISKO evolves both heuristics and designs by systematic variation and selection.
In cooperation with users, EURISKO has designed the winning fleet in two consecutive national-level war game tournaments, and has invented several novel and attractive devices for use in three-dimensional integrated circuit design. It has repeatedly surprised its designer with novel and useful results. Unlike the systems described above, it does learn from experience, though its ability to discover new representations for new knowledge remains sharply limited.
EURISKO and expert systems are called artificial intelligence systems; present CAD systems and silicon compilers are not. What name one uses is unimportant. (I prefer “technical AI systems” to refer to systems capable of flexible automated engineering; this name distinguishes them from the often-discussed “social AI systems” that would talk, listen, and pretend to be human.) What matters to space development is where the technology is going — that is, toward automated engineering of increasing speed and power.
Robots and Space Employment
Advocates of space development and settlement have commonly seen AI systems and robots as threats, as potential competitors to humans in space. This concern appears misplaced. Robotics and AI will let us accomplish many tasks in space with fewer people, but to that same extent they will make many more tasks possible, increasing the overall level of space activity. Before robotics and AI become good enough to replace human workers in space, technical AI systems and industrial robotics will bring a revolution to industry on the ground — including the space industry.
Because space activities are few, space hardware is still produced in small production runs; systems are largely hand-crafted and hence expensive. Further, with few units produced, high design costs add greatly to each unit’s cost. These limits will eventually fall. Technical AI systems will multiply the productivity of human engineers by large factors while speeding their work. Industrial robots, directed by these advanced systems, will drop production times and costs for small quantities of hardware, even as lower costs encourage production of larger quantities.
Advances in AI and robotics will thus cut at the roots of the high cost of space activity, dramatically speeding the pace of space development. By the time that automated systems can replace human labor, they will by the same token be able to build us affordable spacecraft and space settlements. It then will no longer matter whether human beings are “needed” in space — it will suffice that some want to go.
Technical AI systems will speed design, including design of new tools, and new tools will bring dramatic new capabilities. All our present tools (at least, those used outside the biochemist’s laboratory) share a fundamental handicap: they handle atoms not as individuals, but as unruly herds. Rather than arranging atoms according to an engineer’s specifications, processes such as melting, casting, rolling, bending, and cutting merely push masses of atoms about, letting them arrange themselves as unguided thermodynamics and kinetics dictate. We have accomplished much with such methods, but only a tiny fraction of what the laws of nature permit.
Living cells, however, demonstrate that molecular machines can exist, and that some (such as the ribosome) can build other molecular machines with atomic precision. A paper of mine, “Molecular Engineering: an approach to the development of general capabilities for molecular manipulation” (Proceedings of the National Academy of Sciences, Vol. 79, pg. 5275, 1981; see also Smithsonian, November 1982) describes how advances in protein engineering will open a path to the construction of a broad class of molecular devices. This will include assemblers, molecular machines able to position reactive molecular groups and bond them together, building structures to complex, atomic specifications. Review articles on protein engineering have since cited this prospect.
How atoms are arranged makes a difference. The carbon atoms in a lump of coal, rearranged, make diamond. Arranged another way, they will make a strong, fracture-tough diamond-based composite having about fifty times the strength-to-mass ratio of the aluminum used to build the shuttle. Such materials will make possible spacecraft of outstanding performance.
Further, when arranged one way, atoms make up bare, moist dirt. Add sunlight and a small bundle of molecular machinery — a seed — and that dirt becomes covered with crabgrass, producing yet more seeds. Like the molecular machinery of crabgrass seeds, assembler-systems will he able to build copies of themsleves. Under computer control (and with molecular technology, computers will be made remarkably small) assemblers will be able to replicate to large numbers and then cooperate to build macroscopic objects.
This will occur at the end of a long series of developments. The crucial steps include (1) developing the technology base for assemblers and (2) programming assembler-based production systems to make what we want. Advanced technical AI systems will greatly speed these achievements. Assembler technology will yield self-replicating machines able to make products of outstanding quality, including high-performance spacecraft and other space hardware. In space, replicating assemblers will be able to use solar power and asteroidal materials to construct large systems without human labor — systems including space settlements.
The Path to Space
The case for technical AI and molecular technology has barely been outlined here; I treat them in greater depth in a forthcoming book (Engines of Creation, Doubleday, fall 1985). To some these ideas may seem like mere technological optimism, but the engineer’s optimism is the technophobe’s pessimism. In fact, without great care we may not survive the arrival of assembler technology — it will bring unprecedented dangers. But, whether these technologies seem promising or terrifying, competitive pressures virtually guarantee their arrival. Each will emerge through a series of profitable (or strategic) advances. Each will be pursued in the US, Japan, Europe, the Soviet Union, and everywhere else that the institution of research and development has taken root. Only global disaster or global dictatorship could halt the global technology race.
Space advocates have long argued with scoffers who declare space development and settlement to be impractical or impossible goals. We have long replied that these goals demand only vehicles and hardware based on quite conventional technologies — that we can open the space frontier by expanding profitable space activities to achieve economies of scale which will in turn open fresh possibilities. This argument is sound and useful. Because it relies on conventional technology, it offers an effective way to persuade those great masses of people who lack any sense of where technology is going. It is, however, a poor description of the likely course of space development.
We have battled scoffers claiming space development to be impossible, but in fact advanced technology will make it easy. Design and production of space hardware will become cheap, access to space will become affordable, and the door to space will open wide for all who care to use it.
How will the transition occur? At first, government and industry will expand space activities in the more-or-less traditional way. Costs will fall and capabilities will increase. As in the past, each generation of space hardware will at first take about five to ten years to develop. In parallel, advances in CAD, expert systems, and EURISKO-like systems will continue, moving toward advanced systems for automated engineering. Computer hardware and software will advance with a generation time of a few years — again, as they have in the past. Software advances, however, will quicken the pace of space hardware design. This will proceed slowly at first, and then faster. Finally, there will come a time — very likely within the span of a single traditional space-hardware design cycle — when technical AI systems will reach a turning point in ability. Beyond that point, they will accelerate engineering development to many times its earlier pace.
Fairly shortly thereafter, the state of space development will depend little on its state before the technical AI breakthrough. At about the same time — perhaps earlier, perhaps later — technical AI will accelerate molecular technology past the assembler breakthrough, placing space hardware on a new technological foundation. Thus, strangely enough, the timing of large-scale space development may he determined less by advances in modern space hardware than by research in the seemingly unrelated areas of AI and biochemistry.
If we survive the coming upheavals, technical AI and molecular technology will make it as easy to settle space as it was to settle the American west — indeed, far easier. When advanced technology sweeps away the wall around our ancient homeland, we will see that we have been in space all along. Earth will then take its place as simply the oldest part of a far wider human world. To achieve this, our chief challenge is to survive as a free civilization with access to space. Survival itself will demand that we run the technology race swiftly and carefully.
The current space program offers profit, pride, and hope. It stimulates technology development and serves as a vivid reminder that the law of space will be the law of the lands of the future. How we manage conflict and cooperation in space today will affect our chances for simple survival. For all these reasons, what we do in space today will matter to the future of space development, and to the future.
K. Eric Drexler is a member of the L-5 Society Board of Directors and serves as the Associate Editor of the L-5 News. His book Engines of Creation will he published by Doubleday this coming fall.