人工智能和资本主义的未来

The first is Tessa Morris-Suzuki’s examination of Japan’s hightechnology capitalism in the 1980s, the period which saw the introduction of robots in auto manufacture, the explosive growth of the video game industry, and projects such as Japan’s ultimately doomed ‘Fifth Generation Computer Systems’, an early AI initiative supported by Japan’s Ministry of International Trade and Industry. Writing in this context, Morris-Suzuki suggested – much as we have here – that some of the disarray of the contemporary left stems from its reluctance to confront the possibility of a highly automated capitalism, instead taking ‘one of two contrasting positions’ – either denial ‘that the contemporary “information revolution” represents any fundamental change in the nature of capitalism’ or the assertion ‘that it spells the death agony of the capitalist system’ (Morris-Suzuki 1986: 81). Taking issue with the claim by the famous Marxist scholar Ernest Mandel (1975: 207) that large-scale automation of production constitutes the ‘absolute inner limit’ of capitalism, Morris-Suzuki argued it was time to consider ways capital might perpetuate itself under such conditions. These, she said, would include the transfer of labour from production to ‘perpetual innovation’, a proletarianization of technical jobs, the corporatization of an education system geared to the production of elite research scientists, and the creation of a workforce ‘easily taken up and easily discarded’ (Morris-Suzuki 1984: 120). This now reads as a prescient description of the present. The new wave of AI poses the same problem Morris-Suzuki articulated, but at a higher level. ML and other new AI techniques are beginning to encroach on the activities she saw as the only available refuge for labour chased out of industrial production by machines.

The second is Ramin Ramtin’s remarkable Capitalism and Automation: Revolution in Technology and Capitalist Breakdown (1991). In this book, Ramtin made the first systematic attempt to rethink Marx’s theory of machinery in the light of the cybernetic technology that was driving the digital automation of the 1980s. He proposed that to Marx’s three-part anatomization of industrial machinery, comprising motor power, transmission mechanism and tool head, had to be added the guiding or control function – a function once considered dependent on human intelligence and senses but now increasingly automated with information, including sensor, technology. By proposing this revision Ramtin offered a way towards a theorization of computers that, without endorsing the post-industrial euphoria about the information revolution, also recognized the qualitative change digital technologies brought. Ramtin’s work was, however, also notable for the unflinching eye it cast on the possible consequences of this development for a Marxist analysis of class struggle. He suggested that the full-scale cybernetic onslaught of capital against its working class would bring to the fore issues of unemployment that had receded into the background during the postwar boom. His insistence that Marx’s notion of proletarianization be recognized as a concept not just of workplace exploitation but also of the liability of ejection from work in many ways anticipates the discussions of ‘surplus populations’ that would emerge in the wake of the 2008 crash, which we take up later in this book.

22

inhuman power

Morris-Suzuki and Ramtin not only pointed to important changes in work and labour conditions associated with early AI. Between them, they also provided important revisions of Marx’s basic conceptualization of machinery. Ramtin, drawing on cybernetics theory, pointed out that Marx’s account of machines omits the key function of the control, presumably assuming it is ultimately directed by a human agent. This assumption is, however, Ramtin pointed out, precisely what was challenged by cyberneticists as they introduced the theory of feedback into machine operation. It is, he suggested, precisely the control function that distinguishes automation from mechanization and makes it qualitatively new. Morris-Suzuki emphasized the vastly increased scope and flexibility of machine application that comes with the separation of ‘hardware’ and ‘software’: with machinery whose operations can be changed by the switching of instructional programs, so that machines start to attain some of the variability that had previously been seen as unique to human labour.

While Ramtin and Morris-Suzuki provide analytic anticipations of AI drawn from early moments of digital automation, our third exemplar of Marxist AI analysis, George Caffentzis, provides a crucial theorization as to why many of the predictions of jobless futures of that era have not come true. In a series of essays written from 1980 to 2008,

introduction: ai-capital

23

Caffentzis argued that the apparent job-destroying powers of AI had to be considered in light of its antithesis, the expansion of the service sector and global sweatshops; one had to think ‘Africa’ and ‘automata’ together. He draws on the ninth chapter of Volume 3 of Capital, ‘Formation of a General Rate of Profit (Average Rate of Profit) and Transformation of Commodity Values into Prices of Production’ (1991: 254–73). Here Marx suggested that while the profit extracted by capital as a whole depends on the overall amount of surplus-value extracted within its entire system, there is no direct correspondence between any individual capitalist’s profit and the amount of socially necessary labour they employ. Value is a social phenomenon and any and all value produced goes into a social pool after the commodities have been exchanged. But capitalists also appropriate surplus-value from this pool in the form of profit; an individual capitalist will appropriate more profit if their capital, relative to other capitals, is of a larger size, has a higher organic composition, and a higher average profit rate (Marx 1991: 241–73; Caffentzis 2013: 132–4). Thus highly automated businesses syphon-off the surplus-value generated by labour-intensive capital. Caffentzis called this ‘the law of the increasing dispersion of organic composition’, by which ‘every increase in the introduction of science and technology ... in one branch of industry ... will lead to an equivalent increase in the introduction of low organic composition production in [an]other’ (2008: 65). Caffentzis’s account of capital’s contradictory movements towards high-technology (automata) and low-wage-labour (Africa) suggests reflection on how its current AI-fever is induced not only by technological breakthroughs, but by increasing frustrations in finding cheap labour.

AI NOVUM

Before and during the writing of this book, in addition to assimilating materials about Marxist theory and the science, economics and sociology of AI, we read and watched a great deal of AI science fiction (henceforward, AI-SF), including not just the classic films – 2001, Blade Runner, Terminator – that are inevitable points of reference for all AI discussion, but a wave of more recent writings and productions that accompanied the emergence of ML. Some will regard this as evidence of impaired judgement. Harry Collins (2018: 5–13) renames AI as ‘Artifictional Intelligence’ because, he argues, SF has encouraged a widespread overestimate of AI capabilities: depictions of the purported superhuman singularity encourage ‘the surrender’ – human abdication to stupid computer programs. Similarly, Mike Cook (2018), an AI researcher, designates AI-SF, alongside excessive respect for scientists and the over-selling of AI, as one of a set of cultural factors contributing to a ‘basic lack of understanding’ about what ML can and cannot do.

24

inhuman power

We agree, but think that this is not all there is to AI-SF. Darko Suvin (1979) famously proposed that SF cognitively explores potentialities incipient in a society at a given moment by focusing on a ‘novum’ (Latin for ‘new thing’) – a term he borrows from the Marxist theorist Ernst Bloch (1986 [1955]) to denote some new force appearing on the ‘front line of historical process’ (Suvin 1979: 63–84). Centring itself on the novum of AI, and thereby estranging and de-familiarizing our current reality, AI-SF conducts thought experiments about the possible directions – dizzyingly utopian, terrifyingly dystopian or depressingly mundane – of actually-existing AI-capitalism. In the case of AI-SF, the importance of these thought experiments is accentuated because of the now well-documented feedback loop between computer science and science fiction, in which scientists inspired by SF work towards the actualization of its imagined worlds.

It is, of course, quite true that there is not much thought behind many AI-SF thought experiments. Collins is properly scathing about Hollywood representations of AI as a superhuman James Bond villain or sexual manipulator. Such anthropocentric depictions obscure the profoundly ‘inhuman’ nature of AI. Techno-amplifying already familiar cultural industry tropes, they are complicit with corporate promises that AI offers us a future the same as the present, only bigger and better. However, there are also other kinds of AI-SF that are far more critical in their perspective. While composing this book we have sometimes tried to categorize such works according to which of the contending left perspectives – sceptical, accelerationist and abyssal – they might correspond to.

Cyberpunk AI-SF has affinities with the sceptical Marxist perspective on AI, even while it extrapolates technological capacities well beyond what the sceptics would consider plausible. This is because cyberpunk fictions give an unsparing anatomization of capital in an era of intelligent machines. They defamiliarize the squalors of actually-existing AI-capital by projection into a future where truly shining AI fuses with grimy proletarianization to yield a glistening noir. Superbly realized in the original Blade Runner (though not so much in its visually impressive yet oppressively patriarchal Blade Runner 2049 remake), this genre continues to be embellished in recent AI-SF. Judd Trichter’s novel Love in the Age of Mechanical Reproduction (2015) sends its hero on a grotesquely picaresque journey across a decaying Los Angeles where androids and humans coexist in a state of mounting antagonism and illicit liaison, seeking to recover the parts of his disassembled android lover: capital deepens its organic composition while relentlessly maintaining its domination over its fixed and variable components alike. Different in tone and pace, and original in its feminist thematics and domestic settings, yet sharing a similar problematic, is Andromeda Romano-Lax’s Plum Rains (2018). Set in 2029 Japan it deals with a confrontation between a migrant Filipina worker and an intelligent machine over the care of elderly women, opening onto a complex and melancholy meditation on precarity, colonization and slavery. Although works of this sort may end on hopeful prospects of combined android-human revolt or escape, their overall tenor is to emphasize the subjugation of AI to the brutal equation by which capital owns machines, exploits humans, and substitutes one for the other as profit dictates.

introduction: ai-capital

25

A whole other species of AI-SF is, however, a haven of socialist utopianism. The possibilities of AI-based social and ecological planning underpin many of Kim Stanley Robinson’s explorations of postcapitalist futures, most explicitly in his 2312, where one of the forces enabling a break with an old earth ‘decisively under the thumb of late capitalism’ by other planetary settlements is the possibility that ‘the total annual economy of the solar system could be called out on a quantum computer in less than a second’, so that ‘supercomputers and artificial intelligences … make it possible to fully compute a non-market society’ (Robinson 2013: 125). However, the most striking example of this genre is the late Ian Banks’s great series of ‘Culture’ novels, from Consider Phlebas (1988) to The Hydrogen Sonata (2013). In the Culture hyper-intelligent and benevolent ‘Minds’ – evolved AIs – preside over a galactic commonwealth, coexisting with a humanity that, relieved of material need and even mortality by machinic tutelage, continues an adventurous, individualized and complex unfolding of its species-being. Here, in post-scarcity society, planning has been rendered redundant by plenitude – a fictional rendition of ‘fully automated luxury communism’ (Merchant 2015).

Diametrically opposite to such optimistic visions, and closer to our own abyssal perspective, is a series of depictions of AI as a ‘bad novum’ – not because of any anthropocentric enmity, but rather as the systemic culmination of runaway capitalism. The classic example is Charles

26

inhuman power

Stross’s Accelerando (2005), in which what initially seems a light-hearted tale of digital entrepreneurialism turns into a harrowing story of how AIs proceed to dismantle the solar system as raw material for their ever expanding computational, marketized network, eating the universe in a competitive race that casually discards hominids as sub-optimal, under-performing agents. What Accelerando proposes is that, contra the ecstatic visions of Kurzweil and other utopian prophets of the Singularity, the real meaning of such an event is likely to be that ‘The destiny of intelligent tool using life [i]s to be a stepping stone in the evolution of corporate instruments’ (Stross 2005: 240; see Shaviro 2009). A somewhat similar recent vision comes from computer scientist and AI researcher Zachary Mason (who has worked on Amazon’s ML-powered recommendation systems) (Locke 2017). In his Void Star (2017) protagonists in a San Francisco beset by inequality and climate change find their personal lives effectively subsumed and assimilated by the activities of indecipherable superintelligent AIs. These AIs ‘engage with humanity only as byproducts affected by their actions, while they compute otherworldly questions of symbol manipulation’ (Locke 2017). Massive and eerie derangements of personal identity result as a billionaire plutocrat attempts to mobilize such AI instrumentally.

The truly great AI-SF is, however, perhaps that which slips across, plays with and permutates the possibilities we have schematically reviewed. Still unsurpassed in this regard (and of special interest because of the author’s evident deep engagement with Trotskyism) is Ken MacLeod’s extraordinary Fall Revolution quartet (2008; 2009), which, by adopting a ‘branching futures’ conceit, comprehends a series of AI social outcomes ranging from human-controlled AI planned-economies to hive-mind digital absorption to a de-growth abjuration of AI. A rather similar effect is achieved by the sequencing of Peter Watts’s recent Sunflowers short story and novella (2018) cycle. In this sequence, the protagonist’s early ecstatic connection to AI yields to cynicism and then horror as she is recruited in an endless interstellar exploration journey under the direction of a narrow AI – ‘the chimp’ (Watts 2018) – because the corporate directors of this colonizing journey dare not instantiate AGI that might escape their control. For this reason, the cargo includes a cryogenically frozen crew summoned out of sleep at century-long intervals whenever an emergency requiring lateral thinking occurs. Yet despite this function, the humans are, it transpires, also expendable on a strict cost-benefit basis. The most recent story in this series ends with the defeat by AI surveillance powers of an attempted human mutiny. At the close, the heroine, who has made herself complicit with the suppression to forestall the oxygen-starvation death of all her co-workers, reflects that the emergence of an autonomous AI might offer the best possibility for the crew’s escape from enslavement. Yet the overall sombre tone of Watts’s universe raises the obvious question of whether such an intelligence would be an ally or an enemy to its proletarian fabricators.

introduction: ai-capital

27

In such instances – that is to say, at its best, rather than its worst – SF is a machine for thinking, and in the case of AI-SF, a machine for thinking about machine thinking and capitalism. For that reason, in the following chapters we occasionally weave references to AI-SF into our analysis.

CHAPTER SUMMARY

In what follows we draw on the work of Marx and Marxist scholars to make our own assessment of the present state and future prospects of capital’s rendezvous with AI.

Chapter 1 presents a political economic account of the current state of what we term the AI industry. It describes the main protagonists – the giant tech companies in the US, China and elsewhere, and their interaction with both state research programmes and communities of open-source AI developers. It then goes on from current analysis to near future prognostication, suggesting that the ambitions of the great AI corporations point towards the establishment of AI as a new component of ‘the general conditions of production’, as a ubiquitous infrastructure, akin to the railways of the first industrial revolution or the electrical utilities of the second, on which all other forms of commodity production and circulation will come to rely.

Chapter 2 takes up the issue of AI and employment from the perspective of autonomist Marxism’s class composition theory. It argues that AI should be seen as a second wave of a cybernetic offensive waged by digital capital against its working class, a new onslaught occasioned by the 2008 economic crisis. After reviewing some applications of AI within the social factory of advanced capitalism, we review the debate amongst futurists and economists as to whether AI will generate an imminent employment apocalypse or just a continuation of capital’s processes of job destruction and creation. Whether or not AI brings about an immediate jobs crisis, many other aspects of its deployment are likely to exert downward pressure on wages and working conditions, and it is already precipitating an array of social struggles in and beyond the workplace.

28

inhuman power

Chapter 3 challenges the (to some) reassuring assumption that capital could not survive omnipresent AI automation. Taking its orientation from value theory, and assessing the long-term possibilities of AGI, it proposes that the humanist assumptions underpinning this belief no longer hold; Homo sapiens is not necessarily the only possible subject of capitalist proletarianization. If AI approaches or attains the horizon of singularity, the vistas that open up are not therefore those of inevitable capitalist collapse, but rather of the elevation of machine capital as a literally automatic subject autonomous from human beings. Social democratic programmes of ‘full automation now’ may therefore merely be positioning themselves as benign accomplices to this trajectory.

Our Conclusion draws out some political assumptions of the preceding analysis. While uncertainty is inescapable in thinking about AI, socialist strategies for reforming AI-capital by introducing a universal basic income and eco-modern techno-planning fail to confront the depth of the problem AI presents to projects of human emancipation. A communist orientation to AI focuses on transforming the ownership of the means of production so that real choices can be made about the adoption or abandonment of such technologies. The emergence of a new mode of production is, moreover, likely to occur under conditions of extreme social conflict and ecological disaster. In this context, it is not only capitalism that will be inhuman, for the form of the human that emerges, if any does, from the struggle against AI-capital will not be the same as that which entered into it.

introduction: ai-capital

29

Means of Cognition

[T]he Microsoft view is that AI needs to be included – or in Microsoft speak, ‘infused’ – in everything, from a simple word processor to a quantum computer.

James Thompson (2018)

THE NEW ELECTRICITY

In 2016, Andrew Ng, Stanford professor, entrepreneur, former Chief Scientist at Baidu and former head of Google Brain, pronounced AI ‘the new electricity’ and argued: ‘Just as electricity transformed almost everything 100 years ago, today I actually have a hard time thinking of an industry that I don’t think AI will transform in the next several years (Lynch 2017). Ng is not the only one to espouse the notion of AI as a basic utility leading to a new industrial revolution – the idea is implicit in proclamations of a ‘fourth industrial revolution’ issued by capitalist institutions such as the World Economic Forum (Schwab 2017). It has also been explicitly advanced by tech guru Kevin Kelly (2014), who predicts that in the near future we will have a ‘common utility’ of ‘cheap, reliable, industrial-grade digital smartness running behind everything … Like all utilities, AI will be supremely boring, even as it transforms the Internet, the global economy, and civilization. It will enliven inert objects, much as electricity did more than a century ago. Everything that we formerly electrified we will now cognitize.’ Companies such as Viv (n.d.) deploy this idea in their business plans, asserting that with their AI platform ‘intelligence becomes a utility’.

Predictions such as those of Ng and Kelly suggest that AI could become part of what Marx referred to as the ‘general conditions of production’ (Marx 1990: 506; 1993: 530), i.e. the technologies, institutions and practices which form the environment for capitalist production in a given place and time. Marx spoke of infrastructure, which includes the means of communication and transport, as a significant component of the general conditions of production. If AI becomes the new electricity, it will be applied not only as an intensified form of workplace automation, but also as a basis for a deep and extensive infrastructural reorganization of the capitalist economy as such. This ubiquity of AI would mean that it would not take the form of particular tools deployed by individual capitalists, but, like electricity and telecommunications are today, it would be infrastructure – the means of cognition – presupposed by the production processes of any and all capitalist enterprises. As such, it would be a general condition of production. We propose the term ‘means of cognition’ – the AI-equivalent to Marx’s means of communication and transport – but insist that it not be conflated with the post-operaismo notion of ‘cognitive capitalism’ (Moulier-Boutang 2011), for reasons we discuss in the conclusion of this chapter.

1

To make this argument first requires a review of the history of capitalism’s adoption of AI, a survey of some existing and anticipated commercial applications of AI founded on the ML approach, and analysis of the contemporary AI industry. While the basis for accumulation in this industry is a highly advanced techno-scientific commodity, it is, like all capitalist enterprises, governed by compulsions to produce surplus-value, i.e. seek profit, compete, attract investment, control markets and defeat rivals through the formation of oligopolies and monopolies. We draw attention, however, to two structural features of this industry that could contribute to AI becoming a part of the general conditions of production: its lavish support and subsidization by neoliberal nation states eager to foster AI development for economic, administrative and military purposes; and the seemingly anomalous presence of a large and vigorous open-source component to AI research, in which tools and templates are distributed for free and worked on cooperatively, but are nevertheless channelled towards the platforms and priorities of AI oligopolists.

We speculate on how, in the near future, ML-enabled functions of cognition and perception could become ubiquitous via applications ranging from simple chatbots up to smart cities and the Internet of Things (IoT). These examples demonstrate some ways AI could be positioned as a general condition of production. This analysis paints a picture which runs counter to post-operaismo’s humanist reconfiguration of the notion of the ‘general intellect’ (Marx 1993: 706) as referring to the novel capacities of a networked multitude. Contrarily, the possible future of AI as part of the general conditions of production supports Marx’s original formulation of the general intellect as capital’s accumulated machinic capacities, excised from social human labour. While AI development does, for the moment, depend largely on the mining and processing of data drawn from a networked multitude, the aim of such development is to attain a whole new level of automation giving capital unprecedented independence from labour.

means of cognition

31