Ashish Arora [a] and Alfonso Gambardella [b]
The changing technology of technological change: general and abstract knowledge and the division of innovative labour *
Vol. 23, 1994
In the past, most innovations have resulted from empiricist procedures; the outcome of each trial yielding knowledge that could not be readily extended to other contexts. While trial-and-error may remain the primary engine of innovation, developments in many scientific disciplines, along with progress in computational capabilities and instrumentation, are encouraging a new approach to industrial research. Instead of relying purely on trial-and-error, the attempt is also to understand the principles governing the behaviour of objects and structures. The result is that relevant information, whatever its source, can now be cast in frameworks and categories that are more universal. The greater universality makes it possible for the innovation process to be organised in new ways: firms can specialise and focus upon producing new knowledge, and the locus of innovation may be spread across both users and producers. More generally the use of general and abstract knowledge in innovation opens up the possibility for a division of labour in inventive activity - the division of innovative labour. The implications for public policy, especially that on intellectual property rights, are discussed.
In The Wealth of Nations, Smith states that “improvements in machinery... (are sometimes made by) philpsophers and men of speculation, who... are often capable of combining together the powers of the most distant and dissimilar objects” (Smith, 1982, p. 114). In fact, little of the technical progress in industry has stemmed from the ability to relate “distant and dissimilar objects”. Most innovations and productivity improvements have resulted from empiricist procedures based on trial-and-error; the outcome of each trial yielding knowledge that could not be properly extended to other situations and contexts. 
This paper argues that this state of affairs is changing. Developments in many scientific disciplines, along with progress in computational capabilities and instrumentation, are encouraging a new approach to industrial research. Instead of relying purely on trial-and-error to find what may work, the tendency is to attempt to understand the principles governing the behaviour of objects and structures, to ‘observe’ phenomena and test hypotheses with sophisticated instruments, and to simulate processes on computers. This is not to
a. Heinz School, Carnegie Mellon University, 5000 Forbes Avenue, Pittsburgh, PA 15213, USA
b. Istituto di Studi Aziendali, Universitâ di Urbino, Urbino and JEFE, Università Bocconi, Milan, Italy
This is a revised version of a paper prepared for a conference in honour of Nathan Rosenberg, The Role of Technology in Economics, 8th-9th of November, 1992, Stanford University. We have benefitted from helpful comments and suggestion from several people. In particular, we would like to thank Sergio Barabaschi, Sub Eswaran, Suresh Konda, Alma Rizzoni, Nate Rosenberg, Ed Steinmueller, Salvo Torrisi, and Antonello Zanfel. The paper has benefited greatly from the helpful comments of two anonymous referees. The customary disclaimers on errors and inadequacies apply.
1. For instance, Bessemer, who developed the steel making process named after him, did not quite understand why or how the process worked. Therefore, when the process was first adapted in Britain, it failed to work satisfactorily, for the ores used contained phosphorous whilst the Bessemer process required an acidic medium (Mowery and Rosenberg, 1989, p. 29).
suggest that industrial research can now do without physical experiments, or that innovations arise from basic research alone. Amongst the most important contributions of Nathan Rosenberg to our understanding of technological change is that innovations are often initiated by signals received in the course of production or from customers and markets, and are based on fairly tedious and (from a scientific view point) mundane activities. Such activities remain the primary engine of innovation.
However, as we shall argue below, relevant information for innovation, whatever its source, can now be cast in frameworks and categories that are more universal. The greater universality makes it possible for the innovation process to be organised in new ways. The opportunities for firms to specialise and focus upon producing new knowledge are enhanced and the locus of innovation may be spread across both users and producers. More generally, the use of general and abstract knowledge in innovation opens up the possibility for a division of labour in inventive activity - the division of innovative labour.
Before proceeding, it is useful to clarify our terminology. We distinguish between knowledge on the one hand, and concrete information, on the other. By the latter we mean ‘facts’ about products, processes, and markets. Knowledge provides the context within which information is interpreted. We shall also use the expression general and abstract knowledge. By ‘abstract’ we mean the ability to represent phenomena in terms of a limited number of ‘essential’ elements, rather than in terms of their ‘concrete’ features. By ‘general’ we mean knowledge that relates the outcome of a particular experiment to the outcomes of other, more ‘distant’ experiments. 
In the distinction between general and abstract knowledge, and the use of more practical, empiricist procedures, which generate concrete information, the reader may have sensed a parallelism with the distinction between ‘science’ and ‘technology’. We have deliberately avoided the use of the latter terminology. Apart from the contentious nature of the distinction between the two, both science and technology (however one chooses to define them) utilise and produce both general and abstract knowledge and concrete information. Moreover, in both, tacit know-how and skills are important.
For our purpose, the distinction that is important is that the use of science or technology for economic objectives entails that one solve problems that are too complex to be adequately represented only in abstract terms, viz in terms of few essential elements. In order to come up with new products or processes that work satisfactorily (though not necessarily optimally) in practice, one has to delve into the complexity of problems. Put differently, general and abstract knowledge has to be combined with concrete information, because one also has to attend to the ‘details’ that are typically ignored by abstract representations.
Thus, industrial research has had to resort to long and systematic experiments with objects and systems. In aeronautics, for instance, engineers have long used wind tunnels to simulate the flying conditions of aircraft. Wind tunnels have enabled them not only to test whether a particular design ‘works’, but also aided in the search for a better design. Given the lack of a general theory of aerodynamics, such long and costly experiments have been the only reliable way to design aircraft (Mowery and Rosenberg, 1982; Vincenti, 1990). Similarly, drug discovery has required the laboratory syntheses of a great many molecules and systematic trials before finding one that showed potential therapeutic effect (e.g. Gambardella, 1994). Process innovation has relied heavily on trial-and-error. For instance, in the design of large scale, continuous chemical processes, it has been no simple matter to go from the lab or bench scale process to producing tonnage levels. One has lacked a sufficiently general and comprehensive understanding to be reasonably sure that a process that worked well at the scale of a few pounds a day would not prove to be inefficient, or even dangerous and unworkable when
2. Similar distinctions between knowledge and information have been drawn by others (Nelson and Winter, 1982). The terms general and abstract knowledge and concrete information are borrowed from Rullani and Vaccà (1987) and Di Bernardo and Rullani (1990). We are aware of the fact that an epistemologist may find our definitions naive. However, our purpose here is to define a working terminology that captures the essence of the phenomenon we are analysing.
producing a few tonnes a day. Consequently, the process of ‘scale up’ has required extensive experimentation by skilled and experienced chemical engineers (Freeman, 1968; Landau and Rosenberg, 1992).
The common feature of these examples is that each trial or experiment yielded knowledge that was ‘local’ in the sense that what was learnt from each test could not be readily extended to different situations or contexts. The outcome of the experiments depended upon many variables in ways that were not properly understood. In order to be able to generalise information from one set of experiments and to relate it to information produced by other experiments, one needed to be able to comprehend the phenomenon being studied in an abstract manner. Only then, could one begin to sort out the unessential differences between situations from the important differences. For instance, random screening did not provide many clues as to why a particular compound was effective against a certain disease, because researchers could not associate the structure and the properties of the molecule with organic disorders. Neither could one predict with any confidence how the effectiveness would vary with the sex or age or other factors. Hence, one needed to conduct a great many experiments in order to be sufficiently confident about the behaviour of the system for a large range of parameter values.
The use of general and abstract knowledge in industrial research has received a great impetus from advances in three areas: theoretical understanding of problems, instrumentation, and computational capability. The complementarity between these three areas is apparent, and progress in all three areas is together changing the ‘technology of technical change’. 
Not only can researchers test theories more rapidly and effectively using sophisticated instruments and greater computational power, they can also test theories that could not be tested using ‘old’ experimentation technologies (e.g. theories about the behaviour of ‘nano’-structures). (Apart from testing of theories, improved instrumentation can also point towards improvements in the theories themselves.) In turn, advances in instrumentation have benefitted from greater and cheaper computational power. Computers are used to control instruments, record observations, and analyse the observations quickly and accurately. The value of computational capabilities depends on the advances in theoretical understanding as well. The use of computer simulation requires that engineers conceptualise problems in abstract forms. They have to formalise them in a mathematical language, and translate the mathematical model into software language. The ability to formalise problems in abstract terms depends critically upon a good theoretical understanding of the problems themselves.
The point is that the availability of extremely cheap computational power extends the application (and hence, the development) of theoretical knowledge.  The analysis of protein structures illustrates the complementarity nicely. A protein chain of 150 amino acids can give rise to 5150 possible molecular structures, a number that is impossible to investigate even with supercomputers. A recently developed theorem, using the principle of energy minimisation, cut the number of valid alternatives to 1502 (New Scientist, 1992). This is still too large a number of possibilities to handle with ‘pen and paper’, but not so with supercomputers. As the example shows, the value of computational power is higher when combined with a sophisticated theoretical understanding of phenomenon under study, and vice versa.
The analysis of molecular structures and their interactions with other molecules best exemplifies the benefits arising from the combination of advances in theoretical knowledge, computational and simulation capability, and new instruments (e.g. Baker, 1986). Rational design of molecules is gradually replacing random, trial-and-error ex-
3. We are indebted to Nathan Rosenberg for this phrase.
4. In some ways, computer simulation may be thought of as a substitute for theoretical understanding. Because simulation makes systematic exploration far more time- and cost-effective than physical tests, it may be more advantageous to perform extensive trial-and-error rather than to try to understand the problem in its generality. Nonetheless, we suggest that over long, the synergy between more efficient computerised trials and deeper understanding of phenomena will more than offset the substitution effect.
periments with a great many materials to find one or a few with desired properties.
The drug industry is one where the rise of the new approach has been most apparent. This industry has been using computers to design compounds for nearly 10 years (Science, 1992). Growth of scientific understanding in molecular biology and genetic engineering has clarified important aspects of human metabolism and the chemical and biological action of drugs. At the same time, powerful new instruments make it possible to examine the behaviour of proteins and molecules. For instance, cell receptors in the human body have particular geometrical structures, and the drug molecule has to bind to them just as a key fits into a ‘lock. By studying the structure of receptors, scientists can design (typically on computer) a theoretical compound that matches a given receptor site, and is expected to counter a certain pathology. This narrows laboratory research and. clinical tests to families of molecules whose characteristics are consistent with the ‘ideal’ molecule (Gambardella, 1994).
The development of new catalysts is another case in point. Much of the technological progress in the chemical and oil industries depends upon the development of new catalysts. Zeolites, for instance, are molecular sieves that separate different mixtures through selective adsorption. The oil industry has used zeolites as catalysts since the late 1950s. Until the early 1980s, zeolite catalysts were developed through laborious empirical correlations based on conventional solid-state chemistry and chemical engineering. Even though the discovery and initial use of zeolites did not owe a great deal to the use of general and abstract knowledge, in recent years their development has been heavily influenced by advances in chemistry, especially molecular sieve science, and by the development of sophisticated instruments and analytical procedures (such as NMR and X-ray diffraction). 
Zeolites have channels and cavities that filter out substances whose molecular size is smaller than the zeolite pores, which vary in size. As a result, knowledge of the zeolite structure, and of the reactant molecules, makes it possible to apply a rational approach to zeolite-based catalysis. Zeolites also exemplify the wide applicability of basic principles. ZSM-5 is a medium-pore zeolite. It was developed in the 1960s by Mobil to convert liquid methanol to gasoline. At that time, chemical engineers did not know exactly how it worked, and they had not made significant use of it for some years (Financial World, 1989). Deeper understanding of the structure and the catalytic action of ZSM-5 has boosted its use in quite a few processes. For example, selectoforming is a process to convert low octane components into high octane components. Selectoforming is limited by the contamination of the desired aromatic compound by larger paraffin molecules which are not filtered by the small-pore selectoforming catalyst. However, complete removal of both small and large paraffin molecules in the reformat entails some loss of the product which is to be cracked into gas. Research showed that ZSM-5 not only separates small and large paraffin molecules from the final product, but also prevents the loss of product into gas, producing higher octane compounds without affecting gasoline yields. 
New materials is another field where general and abstract knowledge is being applied with good effect. Models which are based on the relationship between molecular structure and the properties of materials are used to guide the search for new materials. Such models can be improved by observing the behaviour of microscopic particles using new powerful instruments, and by modelling microscopic structures on computer. A recent study conducted at the Pacific Northwestern Laboratory (PNL) reports the result of interviews with several R&D managers in this field (Eberhardt et al., 1991). The managers
5. “While the development of new catalysts was empirical fifteen years ago, research innovations in chemical sciences over the latter years are converting catalysis from an art to science” Research Briefings, 1983, p. 79). “Before 1980 catalysts were synthesized and manually tested in bench-scale reactions to achieve a reasonable level of activity, selectivity, and life; subsequently, the synthesis was modified by trial and error to ultimately obtain economic attractive catalytic performance... Indeed chemical and physical technologies have undergone revolutionary change during the last decade. Technology has unquestionably moved to the molecular level - chemical molecular design is becoming a pervasive methodology - . - The catalyst and preliminary process concept are then designed on paper and certain aspects are simulated by computer” (Cusumano, 1992, pp. 5-6).
6. Cusumano (1992) reports other such examples. See also Research Briefings (1983), Chemical Week (1988 and 1989), Chemical Engineering (1989), Financial World (1989).
emphasized that ‘materials by design’ allows a more complete optimisation of a material because its performance can be simulated under a wide variety of operating conditions; it also enables researchers to eliminate inefficient alternatives before conducting expensive experiments and tests. The study also indicated that materials by design could reduce the time of exploratory research vis-a-vis more traditional trial and error experiments from 5 to 2.5 years, while leading to higher quality products.
The new approach is also being applied to the analysis, design and optimisation of complex systems, like production processes or airplanes. The behaviour of complex systems can increasingly be simulated on computers. This allows the exploration of many different designs, far more cheaply than physical experimentation, and to optimise a number of design features before performing the more costly physical tests. 
One recent important advance in this field has been the development of the so-called ‘genetic algorithms’, which already being used to project turbo-jet engines such as those normally used in regular passenger airplanes (Holland, 1992). Each alternative design is defined by a string that identifies different aspects of the engine (e.g. the shape of internal and external walls, or the pressure, speed and turbulence of the air flow in different points within the cylinder). Each set of characteristics produces a certain performance. The computer uses ‘selective adaptation’ procedures to generate progenies of ‘better’ strings.  It discards strings with low performance, and mixes the characteristics of high performance strings. With sufficient computational power, engineers can scan a great many alternative projects to select a smaller set of designs that blend in a satisfactory way a number of desirable features (even though not yet in a ‘globally’ optimal way). 
One could list other technological sectors as well where general and abstract knowledge is being applied, but it is not truly necessary. Although we may not have provided conclusive evidence, we think that our examples are highly suggestive of the way in which the nature of the innovation process is changing. Not only is this change interesting in itself, we believe that it also has important implications for the organisation of innovative activity, and it is to the latter that we now turn.
The thrust of our argument is that the use of generalised knowledge and abstraction increases the proportion of relevant information that is articulable in universal categories. Hence, it makes a greater fraction of information intelligible and applicable in diverse contexts. When innovations depended primarily on trial-and-error procedures based on physical experiments, much of the knowledge base of the firm was experience-based and tacit. The research process that was carried out based on such firm specific knowledge produced information that was ‘local’ and context dependent. Almost by definition, context dependent information could not be used by an agent unfamiliar with the context within which the information was generated (or only at a great cost). This implied that the innovation process worked best when the innovator possessed the downstream complementary assets needed to develop and commercialise the innovation.
As concrete information comes to be related to more general classes of phenomena, it becomes less context dependent, and can be codified in ways that are more meaningful and useful for other firms as well. Furthermore, as more firms utilise general and abstract knowledge, their frameworks for organising and representing in-
7. We would like to thank Sergio Barabaschi for an illuminating discussion of the use of simulation experiments and ‘virtual prototypes’ in industry.
8. Each individual project for a turbojet engine involves more than one hundred variables and about fifty constraints. The need for great computational power is self-evident.
9. A study showed that, using traditional engineering techniques, an engineer needed about 8 weeks to come out with a satisfactory project of a turbo-jet engine. Another engineer, assisted by an expert system as a ‘seed’ for the genetic algorithm, a third engineer took about 2 days to develop a project with three times as many improvements as in the ‘traditional’ engineer case and one and a half times as many improvements as the engineer assisted by the expert system (Holland, 1992).
formation tend to overlap to a greater degree than in the past. The developments in communications technology, itself closely related with advances in computer technologies, also reduce the costs of inter-firm communication.
To be sure, we are not saying that experience-based learning, and tacit skills and capabilities embedded in organisational routines, are no longer important. Not only does the generation of general and abstract knowledge itself depend on tacit skills and capabilities, but, as discussed in Section 2, firms cannot be content just with understanding problems in abstract terms. In order to come up with specific new products or processes, they have to deal with the complexity and idiosyncratic aspects of applying knowledge to concrete problems, a process which relies heavily upon tacit abilities and trial-and-error. Moreover, the decision to invest in general and abstract knowledge is an economic one. It depends on the complexity of the problem, the radicalness of the planned change, and the diversity of sources of information. If one is contemplating only minor modifications in a process, if one has a relatively short term planning horizon, or if the problem is very complex, it would be more sensible to adopt an empiricist approach rather than try to understand the process in fundamental ways.
Our point is that the changing technology of technical change is making the production process of new technologies more divisible. Boundaries between various sub-tasks can be more usefully drawn because the output from the different tasks can be represented in terms of abstract and universal categories, and hence be combined with each other. Within the boundaries, idiosyncratic information, and tacit knowledge and skills would continue to play an important role. In sum, the body of knowledge and information for innovation has become more ‘divisible’ - pieces of knowledge, and bodies of expertise and (tacit) information can be separated into different organizations and re-assembled at a later stage. The use of generalised knowledge and abstraction in industry may thus have important implications for the boundaries of the innovating firm, and more generally for the theory of the firm itself. We suggest that general and abstract knowledge encourages a division of labour in innovation - the ‘division of innovative labour’. 
While many sectors and economic activities have shown a fairly extensive division of labour, innovations has typically been an exception. But what has constrained the division of innovative labour and limited the market for technology? Why is it that the predominant mode of industrial research has been as a part of enterprise which also carries out activities such as manufacturing and distribution? The transaction cost literature provides one perspective. Teece (1988) argues that contracts for intangible outputs are very difficult to specify ex-ante, and problems of lock-in can arise due to sunk costs. Alongside, such contracts may encounter formidable problems of appropriation because of the natural difficulties in appropriating (and hence exchanging) knowledge and information (Nelson, 1959; Arrow, 1962). These factors raise the potential for opportunistic behaviour.
Rather than enter into a debate about the merits of the transaction cost perspective, a more useful approach is to ask the following. What factors determine the costs of writing feasible and efficient contracts which would allow a division of labour in inventive activity? We submit that there is a ‘technical’ constraint upon the division of innovative labour which is logically distinct from the constraint posed by opportunism. This constraint arises from the fact that relevant information for innovation can be strongly context-dependent; more generally, the knowledge base of a firm, which provides the context for interpreting and utilising information, can be highly firm-specific. Consequently, the cost of transferring information across firm boundaries can be far higher than those of intra-firm transfers. Equally important, context dependent information can be contracted for only with great difficulty. However, as firms increasingly use knowledge bases that are more universal categories, and produce information that is usable in a number of different contexts, the costs of contracting decrease. 
As argued by many authors, large and small firms have a ‘natural comparative advantage’ in
10. Wes Cohen and Paul David have joint claims on the paternity of the phrase.
[11. Clearly, there can still be problems of appropriation, and hence a division of innovative labour will have to rely on strong intellectual property rights. See below for further discussion. Also, our division of innovative labour will not be characterised by arms-length market transactions. The need for complementary tacit knowledge, skills and assets, and the need to restrain opportunistic behaviour, imply that it will have to rely on more complicated forms of governance structure, like joint-ventures and collaborative alliances. This point is discussed extensively in the burgeoning literature on innovation networks (e.g. Freeman, 1991).]
HHC: [bracketed] displayed on page 530 of original.
different stages of the innovation process. This debate goes at least as far back as Jewkes, Sayers and Stillerman (1958) who argued that a number of well known innovations did not originate with the firms that are now associated with those innovations. Mueller (1962) showed that most of Du Pont’s major innovations between 1920 and 1950 came from outside sources. He concluded that Du Pont’s comparative advantage was in large scale development and improvement of ‘inventions’ rather than in the inventive process itself.
Arrow (1983) has argued that the organisational flexibility of a small firm and the lower organisational distance amongst its internal units reduce asymmetric information between innovators and people making decisions about internal allocation of resources. Smaller firms then have greater incentives to carry out more novel and riskier innovation projects, provided that they can finance them. Big firms are better at large scale development, production and marketing. Similarly, Holmstrom (1989) has argued that different organisational structures have differential advantages in performing hard to measure activities like innovation vis-a-vis more routine activities (like development and commercialisation). Holmstrom argues that bureaucratisation that charactenses large firms is an optimal organisational response to the need of coordinating many tasks in large firms, but is hostile to innovation.
The empirical evidence on the question of firm size and ‘innovativeness’ is mixed (Cohen, 1992). But if our premise is correct, the mixed evidence should come as no surprise. A division of innovative labour has been severely , impeded by the factors discussed above. Hence, smaller firms, even though in principle more efficient, would be less likely to invest in innovation.  Not only would an innovative small firm be faced with the difficult task of acquiring the necessary downstream assets required for commercialisation, it might also be handicapped by the large fixed costs of creating a firm-specific knowledge base called for by the trial-and-error approach to innovation. If our premise is correct therefore, we should expect to see an increase in the ‘innovativeness’ of small firms in the future. 
Not only does the use of general and abstract knowledge promote a division of innovative labour, the converse is true as well. A more extensive division of labour would imply ‘thicker’ markets for information and knowledge based services. With many sellers and buyers, the uncertainty of having to rely on outside sources would be reduced, and hence also the incentives to vertically integrate. As the markets develop, the incentives to comprehend and articulate the knowledge base of a firm in terms of more universal categories are likely to increase as well since this would enhance the ability to participate in the division of innovative labour. If intellectual property rights adapt, smaller firms can profitably invest in knowledge-based products and services. Correspondingly, larger firms could specialise in large scale development and marketing, and in research based on lumpy assets. A division of innovative labour would then be socially desirable as it would allow firms to specialise in the activities to which they are comparatively better suited. 
12. Cohen and Klepper (1992) provide a related explanation. Larger firms can spread the cost of the innovation over a larger output, and hence would invest more than smaller firms.
13. Not surprisingly, an extensive division of innovative labour between small and large firms is observed today in the pharmaceutical industry. (See, inter alia, Arora and Gambardelta (1993)). Molecular biology and genetic engineering have supplied a generalised method for discovering new drugs. This, along with a strong system of intellectual property rights, has encouraged specialisation of biotech companies in upstream research, and changed the structure of the industry from one where innovation was highly integrated (from research to distribution) in large firms.
14. Arrow (1983) also noted the possibility of a division of labour in innovation according to firm size, although he did not explicitly recognise the constraint posed by the context-dependent nature of technological information: “Availability of research outcomes on the market will reduce the incentives... [of large firms] to use only internally generated research outcomes... There are limits to relying on the market for research inputs... But clearly some substitution will take place.” (Arrow, 1983, pp. 26-27.)
Another way of looking at the division of innovative labour is to examine it in the context of user-producer interactions. Von Hippel (1990) argues that a great deal of information is ‘sticky’, in the sense that it is very costly to transfer across organisations. He suggested that economic benefits can arise if one could shift the locus of problem-solving rather than moving the sticky information. For example, producers of application-specific integrated circuits (ASIC) used to acquire detailed information about the needs of their clients before designing the customized cirçuits. This was inefficient because a lot of this information was costly to transfer. ASIC manufacturers now supply their clients with ‘generic’ components, along with user-friendly CAD-software packages so that users can adapt the basic component to their needs. 
Our conceptualisation advances Von Hippel’s in that we also analyse the factors that enable one to move the locus of problem-solving. In terms of the framework developed here, in order for the locus of problem-solving to move across organisational boundaries, problems must be conceptualised in general and abstract forms. For instance, ASIC manufacturers had to create circuits that were general enough and flexible enough that they could be tailored to a number of different applications. Moreover, the customisation was being done by the users themselves, many of whom might not have any significant expertise in the design of ICs. To make this possible, the diverse applications for ICs had to be conceptualised in a sufficiently general (and hence also, abstract) manner. Equally important, the parameters of IC design had to be related to the parameters of IC use (as understood by users). Only then could users adapt a generic tool to their specific needs. (See also Steinmueller, 1992.)
In our characterisation, furthermore, the boundaries of vertical integration are not given. Sticky, or context-dependent information, as we prefer to call it, encourages vertical integration. For instance, despite current attempts to develop generalised layers of software on which to build applications, software production is still largely idiosyncratic to specific problems. As a consequence, a great deal of software is still produced by users themselves (Steinmueller, 1993).
As noted above, the idea that the problem of appropriating rents from innovation leads to vertical integration in the innovation process is well established. Levin et al. (1987) found that in most industries, other means of appropriating rent, such a secrecy, and first mover advantage, were more important than patents.
Two points should be noted. First, a division of innovative labour requires better defined and vigorously enforced intellectual property rights. Specialisation in producing information or knowledge based services requires that the producer be able to protect unpermitted use of its ‘output’. A loose system of intellectual property rights would then severely undermine the ex-ante incentives of specialised suppliers to innovate (and therefore to exist), as they would not really have many other means to protect their outcomes.
Second, the very forces which encourage a, division of innovative labour also enable patents (and intellectual property rights more generally) to play a more important role in sustaining such a division of labour. While it is tempting to see the effectiveness of patents solely as a creature of patent policy, one must also keep in mind the role of the underlying knowledge base. The effectiveness of patents also depends on the extent to which the new ideas and knowledge can be articulated in terms of universal categories, cheaply and effectively. As knowledge can be expressed in more general and universal categories, the object of a patent, and its scope and applicability, can be defined more precisely. Hence, general and abstract knowledge would also enhance the ability to use patents to protect innovations. 
Our conceptualisation sheds a different light on patent policy issues. Some authors maintain that broad patents, by reducing the number of innovators, would reduce the diversity of approaches, and hence the rate of technological progress. (e.g. Merges and Nelson, 1990). If our
15. Similarly, instead of developing user-specific applications, software manufacturers are trying to produce generic tools (using object-orientated programming), on which users can build their own specialised applications.
[16. Thus the much greater importance of patents in industries such as chemicals and pharmaceuticals is to be understood as resulting partly from the better articulated knowledge base. As our examples have suggested, these are industries where general and abstract knowledge has been used extensively. The R&D statistics are consistent with our claim. The ratio of R&D expenditure on basic and applied research to development is about 0.5 in Chemicals and Allied Products. The corresponding ratios for other “high tech” sectors such as Electrical Machinery, Aeronautics, and Automobiles are of the order of 0.15. (Landau and Rosenberg .)]
HHC: [bracketed] displayed on page 531 of original.
premise is correct, then the opposite conclusion would hold: broader patents, by encouraging innovators that lack the size and downstream capabilities, would increase the rate of technological progress.
Put differently, all else held constant, narrow patents encourage vertical integration in the generation and commercialisation of innovations. Under weak patent regimes, large firms will generate in-house the technologies that they commercialise, if only because of the limited supply of external research outcomes. Broader patents would instead encourage investments in ideas and products embodying more generalised knowledge, by firms that are relatively more efficient at these activities and relatively less efficient at large scale production and commercialisation. Correspondingly, it would stimulate larger firms to specialise in downstream activities, where they have comparative advantage.
Clearly, broad patents will also protect the innovations of larger firms more effectively. But because large firms can already protect their innovations through other assets, the marginal benefit of broader patents will be higher for small firms. Apart from its direct effect on the incentives to innovate, a strong intellectual property rights regime would also enhance the incentives to trade in technology. The latter applies with particular force to the case where tacit know-how must also be transferred. Stronger intellectual property rights increase the efficiency of contracts for the sale of technology, and hence, increase the incentives for firms to specialise in the production of technology. (See Arora (1991) for details.) In sum, to the extent that different firms or different organisational structures have differential efficiency in different stages of the innovation process, there could be social advantages to broader patents.
We have argued that industrial research and innovation increasingly rely on more generalised and abstract knowledge. The ability to represent concrete information in abstract and universal categories allows it be used in a number of locations and organisations, even those that are ‘distant’ from the source. With suitable intellectual property rights, this encourages a division of labour in innovation, with different firms and organisations specialising in the stages of the innovation process where they have a comparative advantage.
The empirical evidence that we can offer may not be conclusive. Nonetheless, we believe it is highly suggestive of the underlying trends. We have attempted to show that if, as we suggest, the technology of technical change is indeed changing, it will have far reaching effects on the organisation of technical change. Thus far, the so-called ‘high tech’ industries, biotechnology, new materials, semi-conductors and software, have shown the greatest extent of specialisation, and an upsurge of network-like arrangements for innovation. These are the sectors where the use of general and abstract knowledge has been the greatest and where intellectual property rights are well defined and better protected. If our arguments are correct, the changing organisation of inventive activity in these sectors is a harbinger of the future.
Arora, A. and A. Gambardella, 1993, Division of innovative labour in biotechnology, Working Paper 93-30, Heinz School, CMU, Pittsburgh, PA.
Arora, A. and A. Gambardella, 1992, New trends in technological change, Revista Internazionale di Sciencze Sociale 3 (July-September), 259-277.
Arora, A., 1991, Licensing tacit knowledge: intellectual property rights and the market for know-how, Working Paper 91-35, Heinz School, CMU, Pittsburgh, PA.
Arrow, K., 1962, Economics of welfare and the allocation of resources for invention, in: The Rate and Direction of Inventive Activity, NBER (Princeton University Press, Princeton, NJ).
Arrow, K., 1983, Innovation in large and small firms, in: J. Ronen (editor), Entrepreneurship (Lexington Books, Lexington).
Baker, W., 1986, The physical sciences as the basis for modern technology, in R. Landau and N. Rosenberg (editors),
The Positive Sum Strategy (National Academy Press, Washington, DC).
Chemical Engineering, 1989, Designer catalysts are all the rage, September, pp. 31-46.
Chemical Week, 1988, Catalysts ‘88: restructuring for technical clout, 29 Jiine, pp. 20-62.
Chemical Week, 1989, Catalysts ‘89, 28 June, pp. 24-40.
Cohen, W., 1992, Empirical studies of innovative activity and performance, Working Paper, CMU, Pittsburgh.
Cohen, W. and S. Klepper, 1992, The anatomy of industry R&D intensity distributions, American Economic Review 82 (4) (September), 773-799.
Cusumano, J., 1992, Creating the future of the chemical industry: catalysis by molecular design, in: J. Thomas and K. Zamarev (editors), Catalysis in the 21st Century (Blackwell Scientific, New York).
Di Bernardo, B., and E. Rullani, 1990, Il Management e le Macchine, 11 Mulino, Bologna.
Eberhardt, J., J. Young, P. Molton and J. Dirks, 1991, Technological Advancements in Instrumentation: Impacts on R&D Productivity, Draft Report (Pacific Northwestern Laboratory, Richmond, WA).
Financial World, 1989, Agents of change, 5 September, pp. 44-46.
Freeman, C., 1968, Chemical process plant: innovation and the World Market, N.I.E.R. 45, 29-51.
Freeman, C., 1991, Networks of innovators: A synthesis of research issues, Research Policy 20, 499-514.
Holland, J., 1992, Genetic algorithms, Scientific American 267 (1), 66-72.
Holmstrom, B., 1989, Agency costs and innovation, Journal of Economic Behavior and Organization 12, 305-327.
Jewkes, J., D. Sawyers and R. Stillerman, 1959, The Sources on Invention (W.W. Norton, New York).
Landau, R. and N. Rosenberg, 1992, Successful commercialization in the chemical process industries, in R. Landau and N. Rosenberg (editors), Technology and the Wealth of Nations (Cambridge University Press, Cambridge, MA).
Levin, R. et a1., 1987, Appropriating the returns from industrial research and development, Brookings Papers on Economic Activity 3, 783-820.
Merges, P., and R. Nelson, 1990, On the complex economics of patent scope, Columbia Law Review 90 (4), 840-916.
Mowery, D. and N. Rosenberg, 1989, Technology and the Pursuit of Economic Growth (Cambridge University Press, Cambridge, UK).
Mueller, W.F., 1962, The origin of the basic inventions underlying Dullont’s major product and process innovations, 1920 to 1950, in R.R. Nelson, ed., The Rate and Direction of Inventive Activity (Princeton University Press, Princeton).
Nelson, R., 1959, The simple economics of basic scientific research, Journal of Political Economy 67, 297-306.
Nelson, R. and S. Winter, 1982, An Evolutionary Theory of Economic Change (Harvard University Press, Cambridge,
New Scientist, 1992, The shape of proteins to come, 9 May, p.16.
Rullani, E., and S. Vacca, 1987, Scienza e tecnologia nello sviluppo industriale, Economia e Politica Industriale 53, 3-41.
Steinmueller, E., 1992, The economics of flexible integrated circuit technology, Review of Industrial Organization 7, 327-349.
Teece, D., 1988, Technological change and the nature of the firm, in: G. Dosi et al. (editors), Technological Change and Economic Theory (Pinter, Lonion).
Vincenti, W., 1990, What Engineers Know and How They Know It (The Johns Hopkins University Press, Baltimore).
Von Hippel, E., 1990, The impact of ‘sticky’ information on innovation and problem-solving, Working Paper No. 3147-90-BPS, April (MIT Sloan School of Management, Cambridge, MA).
Research Briefings, 1983, Opportunities in Chemistry, (National Academy Press, Washington, DC).
Science, 1992, The ascent of odorless chemistry, 17 April, Vol 256, pp. 306-308.
Smith, W., 1991, Molecular mechanisms of aspirin action, drug news & perspectives, Vol. 4 (6), pp. 362-366.