6. Agoric systems in the large

In describing the idea of market-based computation and some of its implications, this paper has implicitly focused on relatively isolated systems of software performing relatively conventional functions. The following examines two broader issues: how market-based computation could interact with existing markets for software, and how it could be relevant to the goal of artificial intelligence.

6.1. Software distribution markets

An agoric open system would provide a computational world in which simple objects can sell services and earn royalties for their creators. This will provide incentives that differ from those of the present world, leading to qualitative differences in software markets.

6.1.1. Charge-per-use markets

Perhaps the central problem we face in all of computer science is how we are to get to the situation where we build on top of the work of others rather than redoing so much of it in a trivially different way.

----------------------R. W. Hamming, 1968 [49]

Consider the current software distribution marketplace. Producers typically earn money by charging for copies of their software (and put up with extensive illegal copying). Occasional users must pay as much for software as intense users. Software priced for intense users is expensive enough to discourage purchase by occasional users-even if their uses would be of substantial value to them. Further, high purchase prices discourage many potentially frequent users from trying the software in the first place. (Simply lowering prices would not be more efficient if this lowers revenues for the sellers: with lower expected revenue, less software would be written, including software for which there is a real demand.)

Now consider trying to build and sell a simple program which uses five sophisticated programs as components. Someone might buy it just to gain access to one of its components. How large a license fee, then, should the owners of those components be expected to charge the builder of this simple program? Enough to make the new program cost at least the sum of the costs of the five component programs. Special arrangements might be made in special circumstances, but at the cost of having people judge and negotiate each case. When one considers the goal of building systems from reusable software components, with complex objects making use of one another's services [50], this tendency to sum costs becomes pathological. The peculiar incentive structure of a charge-per-copy market may have been a greater barrier to achieving Hamming's dream than the more obvious technical hurdles.

In hardware markets, it can be better to charge for the use of a device than to sell a copy of it to the user:

Why was [the first Xerox copier] so successful? Two thing contributed to the breakthrough, McColough says. . .technical superiority. . .and equally important, the marketing genius of the pricing concept of selling [the use of the copier], not machines. `One aspect without the other wouldn't have worked,' he said. `. . .we couldn't sell the machines outright because they would have been too expensive.'

--------------------Jacobson and Hillkirk, 1986 [51]

Agoric systems will naturally support a charge-per-use market for software. In any market, software producers will attempt to extract substantial charges from high-volume users. With charge per use, however, the charges to be paid by high-volume users will no longer stand in the way of low-volume users; as a result, they will use expensive software that they could not afford today. At the same time, high-volume users will experience a finite marginal price for using software, rather than buying it and paying a zero marginal price for using it; they will cut back on some of their marginal, low-value uses. The overall benefit of numerous low-volume users making high-value use of the software will likely outweigh the loss associated with a few high-volume users cutting back on their low-value uses, yielding a net social benefit. It seems likely that some of this benefit will appear as increased revenues to software producers, encouraging increased software production.

In a charge-per-copy market, users face an incentive structure in which they pay nothing to keep using their present software, but must pay a large lump sum if they decide to switch to a competitor. A charge-per-use market will eliminate this artificial barrier to change, encouraging more lively competition among software producers and better adaptation of software to user needs.

By enabling small objects to earn royalties for their creators, charge-per-use markets will encourage the writing, use, and reuse of software components-to do so will finally be pro- fitable. Substantial improvement in programming productivity should result; these improvements will multiply the advantages just described.

6.1.2. Hardware encapsulation

This charge-per-use scenario presents a major technical problem: it depends on the ability to truly protect software from illicit copying. True encapsulation would ensure this, but true encapsulation will require a hardware foundation that blocks physical attacks on security. Two approaches seem feasible: either keeping copies in just a few secure sites and allowing access to their services over a network, or developing a technology for providing users with local secure sites to which software can migrate.

In the limit of zero communication costs (in terms of money, delay, and bandwidth limitations), the disincentive for remote computation would vanish. More generally, lower communication costs will make it more practical for objects located on remote machines to offer services to objects on user machines. Remote machines can provide a hardware basis for secure encapsulation and copy protection-they can be physically secured, in a vault if need be. This approach to security becomes more attractive if software can be partitioned into public-domain front-ends (which engage in high-bandwidth interaction with a user), and proprietary back-ends (which perform sophisticated computations), and if bandwidth requirements between front- and back-ends can be minimized.

One system that might lend itself to this approach is an engineering service [13]. The user's machine would hold software for the representation, editing, and display of hardware designs. The back-end system-perhaps an extensive market ecosystem containing objects of diverse functionality and ownership-would provide computation-intensive numerical modeling of designs, heuristics-applying objects (perhaps resembling expert systems) for suggesting and evaluating modifications, and so forth.

Two disadvantages of separating front- and back-ends in this way are communications cost and response time. If hardware encapsulation can be provided on the local user's machine, however, software can migrate there (in encrypted form) and provide services on-site. Opaque boxes are a possible design for such secure hardware:

Imagine a box containing sensors and electronics able to recognize an attempt to violate the box's integrity [52]. In addition, the box contains a processor, dynamic RAM, and a battery. In this RAM is the private key of the manufacturer's public-key encryption key pair [29]; objects encrypted with the public key can migrate to the box and be decrypted internally. If the box detects an attempt to violate its physical integrity, it wipes the dynamic RAM (physically destructive processes are acceptable), deleting the private key and all other sensitive data. All disk storage is outside the box (fast-enough disk erasure would be too violent), so software and other data must be encrypted when written and decrypted when read. The box is termed opaque because no one can see its contents.

Internally, the opaque box would require encapsulation among software objects. This can be done by using a secure operating system [VI], by using capability hardware [53,54,55], or by demanding that objects be written in a secure programming language and either run under a secure interpreter or compiled by a secure compiler [IV,56]. Among other objects, the box would contain one or more branches of external banks, linked to them from time to time by encrypted communications; these banks would handle royalty payments for use of software.

Will greater hardware cost make opaque boxes uncompetitive for personal computer systems? If the added cost is not too many hundreds of dollars, the benefit-greater software availability-will be far greater, for many users. Opaque boxes can support a charge-per-use market in which copies of software are available for the cost of telecommunications. CD-ROMs full of encrypted software might be sold at a token cost to encourage use.

An intermediate approach becomes attractive if opaque boxes are too expensive for use as personal machines. Applications could be split into front and back-ends as above, but back-ends could run on any available opaque box. These boxes could be located wherever there is sufficient demand, and linked to personal machines via high-bandwidth local networks. People (or software) would find investment in opaque boxes profitable, since their processors would earn revenue. With high enough box-manufacturing costs, this approach merges into the remote-machine scenario; with low enough costs, it merges into the personal-machine scenario.

6.1.3. Inhibiting theft

As society embodies more and more of its knowledge and capabilities in software, the theft of this software becomes a growing danger. An environment that encourages the creation of large, capable, stand-alone applications sold on a charge-per-copy basis magnifies this problem, particularly when the stolen software will be used in places beyond the reach of copyright law.

A charge-per-use environment will reduce this problem. It will encourage the development of software systems that are composites of many proprietary packages, each having its security guarded by its creator. Further, it will encourage the creation of systems that are distributed over many machines. The division and distribution of functions will make the problem faced by a thief less like that of stealing a car and more like that of stealing a railroad. Traditional methods of limiting theft (such as military classification) slow progress and inhibit use; computational markets promise to discourage theft while speeding progress and facilitating use.

6.1.4. Integration with the human market

It has been shown how an agoric system would use price mechanisms to allocate use of hardware resources among objects. This price information will also support improved decisions regarding hardware purchase: if the market price of a resource inside the system is consistently above the price of purchasing more of the resource on the external market, then incremental expansion is advantageous. Indeed, one can envision scenarios in which software objects recognize a need for new hardware, lease room for it, and buy it as an investment.

It has been shown how objects in an agoric system would serve human needs, with human minds judging their success. Similarly, when objects are competent to judge success, they can hire humans to serve their needs-for example, to solve a problem requiring human knowledge or insight.

Conway's law states that "Organizations which design systems are constrained to produce systems which are copies of the communications structures of these organizations" (from [57] as quoted in [58]). If so, then software systems developed in a distributed fashion can be expected to resemble the organization of society as a whole. In a decentralized society coordinated by market mechanisms, agoric systems are a natural result.

6.2. The marketplace of mind

Artificial intelligence is unnecessary for building an agoric open system and achieving the benefits described here. Building such a system may, however, speed progress in artificial intelligence. Feigenbaum's statement, "In the knowledge lies the power", points out that intelligence is knowledge-intensive; the "knowledge acquisition bottleneck" is recognized as a major hindrance to AI. Stefik has observed [VII] that this knowledge is distributed across society; he calls for a "knowledge medium" in which knowledge contributed by many people could be combined to achieve greater overall intelligence.

Agoric systems should form an attractive knowledge medium. In a large, evolving system, where the participants have great but dispersed knowledge, an important principle is: "In the incentive structure lies the power". In particular, the incentives of a distributed, charge-per-use market can widen the knowledge engineering bottleneck by encouraging people to create chunks of knowledge and knowledge-based systems that work together.

Approaches based on directly buying and selling knowledge [VII,23] suffer from the peculiar incentives of a charge-per-copy market. This problem can be avoided by embodying knowledge in objects which sell knowledge-based services, not knowledge itself. In this way, a given piece of knowledge can be kept proprietary for a time, enabling producers to charge users fees that approach the value the users place on it. This provides an incentive for people to make the knowledge available. But in the long run, the knowledge will spread and competition will drive down the price of the related knowledge-based services-approaching the computational cost of providing them.

Agoric open systems can encourage the development of intelligent objects, but there is also a sense in which the systems themselves will become intelligent. Seeing this entails distinguishing between the idea of intelligence and the ideas of individuality, consciousness, and will. Consider the analogous case of human society.

It can be argued that the most intelligent system now known is human society as a whole. This assertion strikes some people as obvious, but others have a strong feeling that society should be considered less intelligent than an individual person. What might be responsible for these conflicting views?

The argument for the stupidity of society often focuses not on the achievements of society, but on its suboptimal structure or its slow rate of structural change. This seems unfair. Human brains are presumably suboptimal, and their basic structure has changed at a glacial pace over the broad time spans of biological evolution, yet no one argues that society is worse-structured than a brain (what would this mean?), or that its basic structure changes more slowly than that of a brain. Great intelligence need not imply optimal structure, and suboptimal structure does not imply stupidity.

Other arguments for the stupidity of society focus on the behavior of committees, or crowds, or electorates. This also seems unfair. Human beings include not only brains but intestines; our intelligence is not to be judged by the behavior of the latter. Not all parts need be intelligent for a system to be so. Yet other arguments focus on things individuals can do that groups cannot, but one might as well argue that Newton was stupid because he did not speak Urdu. A final argument for the stupidity of society focuses on problems that result when a few individuals who are thought to somehow represent society attempt to direct the actions of the vast number of individuals who actually compose society-that is, the problems of central planning, government, and bureaucracy. This statement of the argument seems an adequate refutation of it.

The argument for society's intelligence is simple: people of diverse knowledge and skills, given overall guidance by the incentives of a market system, can accomplish a range of goals which, if accomplished by an individual, would make that individual a super-human supergenius. The computer industry is a small part of society, yet what individual could equal its accomplishments, or the breadth and speed of its ongoing problem-solving ability?

Still, it is legitimate to ask what it means to speak of the "intelligence" of a diverse, distributed system. In considering an individual, one commonly identifies intelligence with the ability to achieve a wide range of goals through complex information processing. But in agoric systems, as in human society, the component entities will in general have diverse goals, and the system as a whole will typically have no goals [59]. Nonetheless, a similar concept of intelligence can be applied to individuals, societies, and computational markets.

Individuals taking intelligence tests are judged by their ability to achieve goals set by a test-giver using time provided for the purpose. Likewise, the intelligence of a society may be judged by its ability to achieve goals set by individuals, using resources provided for the purpose. In either case, the nature and degree of intelligence may be identified with a combination of the range of goals that can be achieved, the speed with which they can be achieved, and the efficiency of the means employed. By this measure, one may associate kinds and degrees of intelligence not only with individuals, but with corporations, with ad-hoc collections of suppliers and subcontractors, and with the markets and institutions that bring such collections together at need. The idea of intelligence may thus be separated from the ideas of individuality, consciousness, and will.

The notion of intelligence emerging from social interactions is familiar in artificial intelligence: Minsky [60] uses the society metaphor in his recent work on thinking and the mind; Kornfeld and Hewitt [61] use the scientific community as a model for programs incorporating due process reasoning. Human societies demonstrate how distributed pieces of knowledge and competence can be integrated into larger, more comprehensive wholes; this process has been a major study of economics [8] and sociology [63]. Because these social processes (unlike those in the brain) involve the sometimes-intelligible interaction of visible, macroscopic entities, they lend themselves to study and imitation. This paper may thus be seen as proposing a form of multi-agent, societal approach to artificial intelligence.