人工智能和资本主义的未来

EDGE, CLOUD AND ECONOMIES OF SCALE

Essential to the AI industry, whether at the level of firm or nation, is the expensive hardware on which AI runs. Today this predominantly takes the form of ‘the cloud’ – vast, energy guzzling data centres that users can pay to access over the internet (Mosco 2014). As the availability of bandwidth and processing power have increased, the cloud has become available not only for storage, database and computing functions, but also in the past few years for cloud AI or ‘AI as a service’ (MSV 2018). AI-infused consumer devices and services send data to the cloud where the actual AI processing is done. The cloud also enables the insertion of ML techniques (like image and voice recognition) into websites or programs, as well as the online building of ML models. The intense computational requirements for training deep ML models mean that few small companies can afford to purchase the requisite hardware and instead buy computing time from cloud providers. Tractica (2018) reports that the AI industry has seen a 300,000 times increase in computing power requirements since 2012, making cloud AI a space of aggressive competition. Amazon Web Services and Microsoft Azure have been the top performers, with Google Cloud, IBM and Baidu Cloud attempting to gain market share. Cloud AI thus promises to make ML available to companies well beyond the inner circles of the AI industry, but already ownership of the cloud is consolidated in the hands of a ‘very select few’, with the tech giants being the dominant providers (Miller 2018).

The cloud is complemented by an emerging technique called edge computing. Edge computing is an approach in which some processing is done locally on devices rather than sent to the cloud. Tractica (2018) estimates that the amount of AI edge devices shipped will increase from ‘161.4 million units in 2018 to 2.6 billion units worldwide annually by 2025’. Edge computing offers advantages over the cloud when it comes to bandwidth, network latency issues and security. In applications such as autonomous vehicles, a bit of lag could be disastrous. In addition, keeping data on board a device offers obvious security benefits against hackers. As computational requirements and numbers of users grow, edge computing also becomes attractive to cloud AI providers in so far as it can reduce the load on their clouds (Miller 2018). No one, however, expects the edge to replace the cloud. The two work in tandem. Rather than a decentralized disruptor of the cloud, the edge is likely to become another axis of competition for the tech giants, who are already invested heavily in it.

42

inhuman power

In sum, the AI industry is generating many interconnected commercial ventures – an ‘ecosystem’, to use a term widely adopted by those who like to naturalize capitalist activity. AI is beginning to be generalized beyond the tech industry, propelled by wide applicability, financial incentives and technological advancements. The prospects for further generalization indicate to us the plausibility of a future where capital presupposes access to these means of cognition, or, in other words, where AI becomes part of the general conditions of production.

While there may be dramatic changes to the AI industry landscape over the course of this generalization, for now the tech giants occupy an apex position. Control of cloud computing facilities, ownership of large data sets, and the wealth to hire the best from a limited pool of AI talent are some of their many advantages. These factors suggest that the domination of the information technology sector by a handful of corporate behemoths will continue in the age of AI. One likely trajectory for the AI industry is thus that which Marx described as inherent to the general law of capitalist accumulation, namely the centralization and concentration of capitalist power (1990: 777). This is acknowledged even by unequivocally pro-market observers. As The Economist (2017a) argues: ‘It seems likely that the incumbent tech groups will capture many of AI’s gains, given their wealth of data, computing power, smart algorithms and human talent, not to mention a head start on investing.’ And a business-oriented study enthusiastic about the commercial prospects of ML unabashedly acknowledges that there may be a ‘tradeoff between innovation and competition’:

Like most software-related technologies, AI has scale economies. Furthermore, AI tools are often characterized by some degree of increasing returns: better prediction accuracy leads to more users, more users generate more data, and more data leads to better prediction accuracy. Businesses have greater incentives to build [AI] if they have more control, but along with scale economies, this may lead to monopolization. Faster innovation may benefit society from a short-term perspective but may not be optimal in the longer-term. (Agrawal, Gans and Goldfarb 2018: 23)

means of cognition

43

AI BUBBLE?

The latter half of the 2010s saw increasingly widespread speculation on the emergence of an ‘AI bubble’ (Press 2018). In a 2017 report, the business consultancy Gartner gave warning: ‘As AI accelerates up the Hype Cycle, ... most vendors are focused on the goal of simply building and marketing an AI-based product rather than first identifying needs, potential uses and the business value to customers’ (Hernandez 2017). It cautioned its business readers not to fall for ‘AI washing’ or exaggerated claims about the capabilities of AI systems. More luridly, a 2018 business report predicted that while AI would eventually ‘mature’ as a commercial sector, ‘the field will be littered with corpses on the way’ (Riot Research 2018). Might the AI industry descend into another ‘AI winter’ before the technology is sufficiently generalized so as to constitute a part of the general conditions of production?

The bursting of an ‘AI bubble’ and the failure of many of the recent entrants into the AI industry would not necessarily spell the end of AI-capital. The collapse of overvalued start-ups, and even large, established firms, making room for fresh entrants learning from their predecessors’ failures, has for centuries been a normal cyclical feature of capital’s innovations, from the railway and telegraph onward. This is all part and parcel of capital accumulation. The most germane recent example is the bursting of the US ‘dot.com’ bubble in 2000, which laid waste to many early e-commerce attempts and, by knock-on effect, threw the telecommunications industry into crisis. For a few years this disaster stalled investment in new digital business, until the emergence of Google, Facebook and other Web 2.0 companies, who, with revised business models, joined the big survivors of the bust, such as Amazon and Microsoft, in a new and even larger wave of digital commodification. A similar crisis could be a mere bump in the road along which AI advances.

There are, however, other possibilities that might more deeply disturb booming forecasts for the AI industry. One would be the discovery of serious, endemic problems in the technology itself. Roman V. Yampolskiy (n.d.), a researcher associated with the AI Safety movement, has compiled a short inventory of failures in AI. Perhaps the most serious of these dates from 2013: ‘Object recognition neural networks saw phantom objects in particular noise images.’ This refers to a well-documented tendency of ML object recognition systems to confuse, but report with very high confidence, certain abstract patterns with real-life entities such as peacocks or leopards. This bizarre quirk appears to derive from such systems’ interpretation of radically new objects as merely extreme examples of the data sets on which they have been trained. The most immediate concern arising from this is that image-recognition systems used for military and security purposes are susceptible to malign spoofing by ‘adversarial images’ (Nguyen, Yosinski and Clune 2015). Even more troubling is the implication that neural networks are opaque to their producers, and can suddenly throw up entirely unanticipated results (Scharre 2018). Such disquiet has yet to seriously dampen AI enthusiasm. But disastrous AI errors might dent business confidence. A series of fatal autonomous vehicle accidents in 2018 cast a pall, however temporary, over self-driving car research by reminding corporations and publics of the complex and costly liabilities such technology involves.

44

inhuman power

There is also the possibility that AI may not deliver the goods that capital expects. The entire digital ‘information revolution’ has been bedevilled by the ‘productivity paradox’, a problem summed up in the sardonic observation by Robert Solow (1987) that ‘You can see the computer age everywhere but in the productivity statistics.’ Apart from a brief period at the end of the 1990s, widespread adoption of digital technologies in advanced capitalist economies has not (as of 2019) generated observable year over year productivity increases comparable with those yielded by earlier cycles of innovation. The reasons are hotly disputed. Some economists argue that this apparent anomaly reflects problems in measuring the real significance of digital activity, or that economic gains of digitization merely need more time to manifest (Brynjolfsson, Rock and Syverson 2017). Others, such as Robert Gordon (2016), insist that the economic consequences of information technology are simply much less than those of inventions such as the automobile, electricity, urban sanitation chemicals and pharmaceuticals that powered US economic growth from 1870 to 1970.

In the uneasy aftermath of the Wall Street crash of 2008, US corporate spending on new means of production, such as equipment and buildings, has been low, especially relative to the amount spent on financial and speculative activities such as dividends, stock buybacks and takeovers (Henwood 2018a). Michael Roberts (2018) proposes that ‘[p]roductivity growth in all the major capitalist economies has slowed because of the failure of capitalists in most economies to step up investment in new technologies’. This picture could be changed, either by dramatic falls in the cost of AI-related technologies, by massive state subsidization of AI infrastructures, or by the threat of wage demands emerging from gradually tightening post-recession labour markets, which would incentivize automation. But there is as yet no absolute guarantee that the current wave of new AI research will actually make it out of the laboratory and into a wide transformation of business practices: the AI revolution might subside with a digitally voiced whimper.

means of cognition

45

THE GENERAL CONDITIONS OF PRODUCTION

Let us suppose that capital’s current love affair with AI is not broken up by performance failures and commitment nerves. What might be the eventual offspring of this union? To elaborate the possible long-term significance of the second era of the AI industry we turn to an often-overlooked Marxian concept: the general conditions of production. What does Marx mean by this category that is mentioned in passing in Capital (1990: 472–3, 474, 505–6, 579, 652), but discussed in some depth in Grundrisse (1993: 308, 524–33, 725)?

To begin, it is helpful to understand what it means that these conditions are general. In Marx’s system, ‘general’ is almost always opposed to ‘particular’, and so it is with the general conditions of production, which he distinguished from the ‘particular conditions of production for one capitalist or another’ that are bought or produced directly by individual capitals to keep production going, and include material inputs (raw materials and intermediate goods), means of production, and labour (Marx 1993: 531).⁵ Whereas particular conditions concern the production of this or that individual capital, the general conditions of production are common to all capitals.

In Grundrisse, Marx spelled out the relationship between an individual capital and the general conditions as ‘a specific relation of capital to the communal, general conditions of social production, as distinct from the conditions of a particular capital and its particular production process (Marx 1993: 533). The general conditions are, therefore, ‘something that benefits (or impedes) all particular capitalist production processes’ (Kjøsen 2016: 65). Infrastructure is illustrative of the nature of the general conditions, because roads, canals or railways ‘benefit not just a single capital, but all individual capitals in a given area’ (Kjøsen 2016: 65). With the general conditions of production, Marx thus described the general milieu in which an individual capital finds itself at a particular historical moment; it is the terrain of both class struggle and capitalist competition.

46

inhuman power

Importantly, these conditions are general because they are potentially available to all individual capitals; this does not mean they are free or practically obtainable by all. The case is different for different conditions. Something might be a commodity as well as a general condition. For example, transportation and communication – e.g. container shipping or the hardware required to connect to the internet – are something that all individual capitals have to pay for and are necessary for contemporary capitalist production. As long as a mode of transportation is common carriage, it counts as a general condition of production. Other conditions are free or are paid for indirectly, through taxation. For instance, any and all capitals find themselves on the world market whether they like it or not. On the other hand, the protection of shipping lanes and the maintenance of transportation infrastructure are paid for by taxes.

Infrastructure is but one vital component to the general conditions of production. As one of us has pointed out, the general conditions include a bewildering array of things: the means of communication and transport; the general use of buildings for production and storage; the market, i.e. the sphere and process of circulation; the political world order; the general state of science, technology and engineering; and, confusingly, also specific kinds of production – such as the production of machinery by machinery and the degree of automation in production – as well as the mass and velocity at which production occurs (Kjøsen 2016: 64).

What is the specific function of the general conditions of production? Marx argued, while discussing infrastructure, that ‘All general conditions of production ... facilitate circulation or ... make it possible ... or ... increase the force of production’ (1993: 530–1). A well-designed system of highways, bridges and tunnels will, of course, facilitate circulation in the sense of making it faster or, if built where there previously were no roads, would make circulation there possible. Speeding circulation makes it possible for capital to accelerate the cycle of extracting surplus-value in production and realizing it in the market (Marx 1992: 203). Similar considerations apply to non-infrastructural general conditions of production; for example, increasing the degree of automation in production intensifies surplus-value extraction because machinery allows the application of physical force far exceeding that possessed by human labour alone (Marx 1990: 509).

means of cognition

47

To really appreciate Marx’s argument about the function of the general conditions of production and thus how AI could function therein, it is necessary to discuss how they develop in lock-step with the mode of production itself. That is, the general conditions always refer to a specific locale and time period, meaning that the general conditions for the period of manufacture were different from the period of large-scale industry, which in turn differed from that of Fordism and so on, all the way up to a possible future AI-capitalism (Marx 1990: 505–6; Kjøsen 2016: 65–6). Importantly, the general conditions for one period may be ‘inadequate’ or ‘unbearable fetters’ to the following one, and will thus have to be adapted or updated so that they become ‘appropriate’ to the new period; when they are appropriate, ‘the mode of production [i.e. period] acquires an elasticity, a capacity for sudden extension by leaps and bounds’ (Marx 1990: 579; Kjøsen 2016: 65–6). That is, the velocity and volume of production can now occur at higher levels than during the previous period.

How are the general conditions adapted to emergent modes of production? To sketch the possible capitalist future of AI we need to elaborate this dynamic. The connection between the general conditions and the mode of production starts at the level of branches of industrial capital (Kjøsen 2016: 66). A revolution in production, for example through the invention of a machine, may force transformations in other branches that ‘are connected together by being separate phases of a process, and yet isolated by the social division of labour, in such a way that each of them produces an independent commodity’ (Marx 1990: 505). Any changes that cause productivity increases in terms of volume and/or speed in one branch require connected branches to adapt in order for the original branch to maintain its new level of productivity. To illustrate this dynamic, Marx referred to the mechanical revolution in cotton-spinning that ‘called forth the invention of the gin’ because without this technology, the supply of cotton would not be able to keep up with mechanized cotton-spinning (1990: 505). With the general conditions, this dynamic is writ large; paying particular attention to the means of communication and transport, Marx argues that it was specifically the generalization of production with machinery and its resultant increase in speed and output that forced a change in the general conditions to become appropriate to large-scale industrial production. Thus, when particular branches of production are closely connected, ‘a revolution in terms of knowledge, technology and organization in one branch propagates throughout related branches, leading not only to growth in productivity, but also increased output, which in turn leads to new chain reactions throughout related branches of production and eventually to a revolution in the mode of production’, which thus becomes elastic in its productive capacity (Kjøsen 2016: 66–7).

48

inhuman power

In addition to large-scale industry requiring an ‘immense transformation’ in transportation and communications networks, it also required that machines could produce large quantities of uniform, precisely tooled machine parts: ‘Large-scale industry therefore had to take over the machine itself, its own characteristic instrument of production, and to produce machines by means of machines. It was not till it did this that it could create for itself an adequate technical foundation and stand on its own feet’ (Marx 1990: 506). Only when machines started producing parts for machines could machinery as such become a general condition of production, meaning it was available to all individual capitals, in adequate quantity and quality. This is not to say that every individual capital must adopt a particular new technology for it to be considered part of the general conditions; it is sufficient that individual capitals have access to it on a more or less equal footing. Competition will compel adoption; it will become part of the general conditions once it is widely used by a critical mass of individual capitals. This may be encouraged or even enabled by states, as for example in the 1990s when the US government fostered business adoption of the internet as an ‘Information Highway’ for commodity circulation (Schiller 1999). The high level of governmental interest in boosting national AI capacities suggests that states are already pushing for another revolution in the general conditions of production. Yet before we can discuss what this push might be a response to, or what form it might take, we need to outline the general conditions of production as they stand today.

THE GENERAL CONDITIONS OF PRODUCTION FOR TWENTY-FIRST-CENTURY CAPITALISM

Marx sketched out how the general conditions of production for the period of manufacture were transformed into those adequate for large-scale industrial production. Since Marx’s time, the capitalist mode of production has gone through at least two other notable periods: Fordism, defined by Taylorism and the assembly line, and the period that followed it, defined primarily by ICTs and logistics, which is most commonly described as post-Fordism (Hardt and Negri 2001). The lack of consensus around how to define post-Fordism is evident in the panoply of names that have been given to the same period: digital capitalism (Schiller 1999), logistical or supply-chain capitalism (Toscano 2011; Cowen 2014; Kjøsen 2016), and cognitive capitalism (Moulier-Boutang 2011). We recognize the amorphous quality of the term post-Fordism, and agree with many critiques made of it (see Amin 1994 for an early overview); and we particularly note that many theorists of post-Fordism have underplayed the continuities and overplayed the qualitative differences that this period has with Fordism. That being said, we also hold that it is evident that substantial changes have occurred since Fordism, such that the designation of a new period is worthwhile. Further, we suggest that we are perhaps entering a new period of the capitalist mode of production, beyond post-Fordism, which we refer to as actually-existing AI-capitalism. This, we suggest, may be seen as a middle phase of a larger mode of cybernetic capitalism (Robins and Webster 1988; Peters, Britiz and Bulut 2009; Tiqqun 2001) which tends towards fully developed AI-capitalism.

means of cognition

49

The accompanying table lays out a schematic history of capitalist modes of production and their attendant general conditions of production. Like any chart which purports to grasp complex phenomena, it oversimplifies and runs the risk of appearing more rigid that it is intended to. The periods listed here overlap with one another and were and still are being passed through in different ways and at different speeds in different places around the world.

Following Marx’s theorization, we propose that a revolution in one branch of the economy could provoke the necessity of widespread AI adoption such that it becomes part of the general conditions. In Chapter 2, we suggest that a large part of the excitement about AI development responds to crises encountered by capital’s globalizing search for cheap labour, and in Chapter 3 we argue that the prospect that AI may overcome the fetters to an advanced form of capital left over from a previous period – the persistent fetter of human labour – needs to be taken seriously. But here we ask how generally available AI might be incorporated into capital’s inherently revolutionary dynamics. It is not difficult to understand why capital must rely, and how it relies, on physical infrastructures, such as roads and ports, to overcome the fetter of space. But what fetters to capital’s valorization might generally available AI overcome? And how might AI become generally available?

50

inhuman power