人工智能和资本主义的未来

One reaction to the problems of GOFAI is the ‘situated, embodied, dynamical (SED) framework’ (Beer 2014: 128); Rodney Brooks, a pioneer in the field, called his approach ‘nouvelle AI’ to emphasize its qualitative break with GOFAI (Copeland 2000). SED researchers are often motivated by ‘Moravec’s paradox’ – the observation by roboticist Hans Moravec that ‘it is comparatively easy to make computers exhibit adult level performance on intelligence tests or playing checkers, and difficult or impossible to give them the skills of a one-year-old when it comes to perception and mobility’ (1988: 15). SED approaches to AI emphasize the irreducible importance of the body – with its perceptual apparatuses and morphology – to cognition: for this school, it is through solving material problems that machines can evolve intelligent behaviours. Such approaches are therefore often concerned with robotics and artificial life in addition to AI. SED practitioners initially focused on very simple, insect-like robots, but in the 2010s more complex, partially humanoid robots became possible and have been introduced into industrial settings. It is possible that some variety of the SED framework could be the next dominant AI paradigm.

Another reaction to GOFAI was machine learning (ML). The ML school, formerly called connectionism, existed as early as the Dartmouth workshop, gained some traction in the 1980s with advances in learning algorithms, but did not explode until the 2010s when big data and cheap computing power proliferated. As of 2019, ML is the dominant approach to AI.³ It is a statistical pattern-recognition approach. One NVidia researcher has described ML as a process comprised of three steps: ‘(1) take some data, (2) train a model on that data, and (3) use the trained model to make predictions on new data’ (Dettmers 2015). In other words, ML systems can be understood as creating their own models of inference.

12

inhuman power

While ML may operate on a variety of architectures, the cutting edge of AI in the 2010s has largely run on artificial neural networks (ANNs) – computer programs that are inspired by, albeit quite different from, the human brain. ANNs roughly mimic the electrical operations of the brain’s neuronal connections rather than emulate high-level logic like GOFAI does. ANNs are ‘based on the assumption that cognition emerges through the interactions of a large number of simple processing elements or units (i.e., ‘neurons’)’ (Sun 2014: 109). Artificial neurons are organized into a series of layers and each layer is connected to the layers above and below. The lowest level receives inputs – e.g. images, text or speech – in the form of data that has been vectorised (converted into long strings of numbers). Higher levels – called hidden layers – process data that is sent up from the layers below them. Early networks had only one hidden layer, but today’s networks have many more. In general, the more layers the underlying ANN architecture of the ML system has, the more complex patterns it can find. The most advanced ML in the 2010s is called ‘deep learning’ because these networks are many layers deep (LeCun, Bengio and Hinton 2015: 436–44), with some networks possessing as many as 1,000 (He et al. 2016). While previous ML networks were constructed by hand, a ‘key aspect of deep learning is that these layers of features are not designed by human engineers: they are learned from data using a general-purpose learning procedure’ (LeCun, Bengio and Hinton 2015: 436).

The artificial synapses which connect the layers of artificial neurons are ‘weighted’ with numeric values representing the strength of the connection. The network ‘learns’ through adjusting the weights of these connections. We can thus think of an ML system as a fixed template with changeable parameters; ‘by assigning different values to these parameters, the program can do different things’ (Alpaydin 2016: 24). ANNs are exposed to a data set, which might be images of faces or audio clips of people saying hello. In a process called ‘training’, the network is exposed to many instances of the chosen object(s) and the weights of the synapses are adjusted by a learning algorithm until the network learns to output the correct response, recognizing faces or the word hello, as the case may be (Kaplan 2016: 30).

introduction: ai-capital

13

It is useful to distinguish between the three broad types of machine learning – supervised, unsupervised and reinforcement learning. Supervised learning has been the most successful so far. In this approach, input data is labelled by human teachers, usually in terms of categories, and the system learns those categories by discerning patterns across the supplied examples. Given enough photos of red hexagonal signs with the word STOP on them, in various visibility conditions, and from various angles, a supervised learning system can learn the concept of a stop sign.

However, because of the necessity of labelling, supervised learning entails a lot of human labour. This has driven companies to develop techniques of unsupervised learning to enable a network to generate categories and labels on its own (Alpaydin 2016: 117). The idea is that with exposure to enough data, the system will identify ‘incredibly sophisticated and complex correlations’ across the data set (Kaplan 2016: 30). In so doing, it may be said to generate its own concept of a stop sign. Some deep learning pioneers argue this will eventually become the central ML approach because it mimics how humans and animals evolved to learn; not by being told what everything in the world is called, but through observing it (LeCun, Bengio and Hinton 2015: 442).

Reinforcement learning lies somewhere in between supervised and unsupervised learning. The pioneers of this approach describe it as ‘learning what to do … so as to maximize a numerical reward signal. The learner is not told which actions to take … but instead must discover which actions yield the most reward by trying them’ (Sutton and Barto 1998: 127). Reinforcement learning was thought to be limited to simple domains, until 2013 when the UK firm DeepMind combined it with unsupervised learning to teach a system to play Atari games with superhuman skill, without programming any knowledge about the games into the system, and giving it access only to the score of the games and the pixel information displayed on the screen (Knight 2017). The same combination enabled AlphaGo’s win over Go master and world champion Lee Sedol.

As Kaplan emphasizes, the learning in ML should be understood as ‘extract[ing] patterns from data’ (2016: 27).⁴ Instead of being built top-down as a set of rules for handling data, ML systems go bottom-up: ‘learning algorithms … are algorithms that make other algorithms … computers [that] write their own programs, so we don’t have to … [it is] the inverse of programming’ (Domingos 2015: 6–7). This is why ML advocates see it as a Copernican revolution in programming. If a system can learn, the designer does not need to anticipate and program a solution for all potential situations the system is exposed to (Alpaydin 2016: 17). Ideally, the system will develop its own solutions and ‘the data itself … defines what to do next’ (Alpaydin 2016: 11).

14

inhuman power

This would not, however, be a very powerful attribute of ML if these systems could not generalize what they have learned to data not included in the training data set. Thus they are assessed according to their ‘generalization ability’ (Alpaydin 2016: 40). While the ML systems of 2019 are at best able to generalize to new data of a similar kind to that which they were trained on, increasingly sophisticated systems may approach AGI’s hypothetical capacities for generalization across domains. This book is about the implications of AI, from machine learning to AGI, for the future of capitalism.

MACHINE MARX

Machinery is crucial in Marx’s analysis of capital, so much so that almost any serious study of his work engages the topic in some way. Here we rapidly map those aspects of Marx’s thought that are most important to the discussion of AI, highlighting points relevant for our arguments later in this book. There are, we suggest, three main strands in Marx’s thoughts about machines. In the first, major line, the machine is a supplement to the human labour that is the crucial creator of value within the capitalist system. While production becomes increasingly machinic and intensifies the exploitation of workers, machinery ultimately contributes to the system’s terminal crisis. The other two, minor lines, logically emerge from this major current but also depart from it. Both of them posit a moment at which the machine becomes autonomous from labour. However, one sees the consequences as liberatory, the other as nightmarish. These three strands of Marx’s machine-thinking can, with care, be conjugated together, but this depends on glossing tensions between them that widen and deepen as we confront the conundrums of AI.

What we term the major line of Marx’s machine analysis unfolds in the first volume Capital. In this account, machines, along with other equipment, buildings and raw materials, are ‘constant, fixed’ capital.⁵ This is contrasted with the ‘variable’ capital of human labour (Marx 1990: 508–9; 1992: 237–48). This distinction between ‘fixed’ and ‘variable’ capital rests on the basic proposition that it is only human labour that creates value within capitalism: the machine, however gargantuan its powers seem relative to those of humans, can only act as a supplement or force-amplifier to the essential, human activity, increasing its efficiency, albeit by manifold times. Machinery, which has itself been built by humans, is ‘dead labour’.

introduction: ai-capital

15

Rather than generating new value, the machine already has value which it transfers to the product: it ‘yield[s] up its own value to the product it serves to beget’ (Marx 1990: 509). The social function of fixed capital is to produce relative surplus-value, which it does by reducing necessary labour-time and, hence, increasing surplus labour-time, i.e. by ‘shortening the part of the working day in which the worker works for himself, to lengthen the other part … he gives to the capitalist for nothing’ (Marx 1990: 492). Increasing the productivity of labour means that the worker’s output is increased: more commodities are produced in less time, and consequently these commodities are cheapened because less value is objectified in each individual commodity.

Marx’s detailed explanation of machinery occurs in Chapter 15 of Capital, Volume 1, which concerns ‘Machinery and Large-Scale Industry’ (1990: 492–639). Here, Marx was concerned ‘only with broad and general characteristics’ of machinery (1990: 492). The genealogy of machinery is found in tools; the instruments of handicraft labour are turned into machinery, automation technology being ‘fully developed machinery’, which has three different parts: (1) the motor mechanism that is the driving force or motive power; (2) the transmitting mechanism that regulates, changes and distributes motion; and (3) the tool or working machine that uses this motion to modify the object of labour (1990: 494). A key component of any machine is its ‘emancipation from the restraints of human strength’ which occurs when it obtains a regular and controllable motive power; Watt’s double-acting steam engine being the first major ‘self-acting prime mover’ (1990: 502). Whereas with mere tools the process of production is adapted to the worker, the system of machinery is a ‘vast automaton’ which confronts the worker as a ‘pre-existing material condition of production’ (1990: 508).

The section on ‘The Struggle Between Worker and Machine’ in Chapter 15 emphasizes how capital’s technological dynamic is inseparable from class conflict. Marx examined the paradox by which, under capital, labour-reducing machinery creates a hell for labourers: the mechanical lightening of demands for physical strength catalyses the large-scale induction of women and children into factories; the capacity of machines to run indefinitely, and the need to pay for their purchase, leads to a prolongation of the working day; the ability to accelerate and multiply machine operations results in the intensification of work. Marx also addressed technological unemployment, describing how ‘[t]he instrument of labor strikes down the laborer’ in a process where machinery ‘act[s] as a superior competitor to the worker, always on the point of making him superfluous, and capital proclaims this fact loudly and deliberately, as well as making use of it’ (1990: 562). Indeed, Marx suggested ‘it would be possible to write a whole history of the innovations made since 1830 for the sole purpose of providing capital with weapons against working class revolt’ (1990: 563).

16

inhuman power

Mechanization is also propelled by competition between rival capitalists. The value of a commodity depends on the amount of socially necessary labour expended in its production. If a capitalist can, by introducing technology, reduce the labour for which she pays, while still selling the product at the prevailing price, she will enjoy greater profits than her competitors. This advantage will eventually be neutralized as use of the labour-saving innovation becomes generalized, but this just sets the scene for the next wave of automation. Thus both class conflict and competition between enterprises give capital an intrinsic drive to replace humans with machines. Marx described this as the tendency for capital to increase its ‘organic composition’, that is to say, the proportion of constant (machines, buildings and raw materials) compared to variable capital (labour) (1990: 762).

Machinery also, however, throws capital into crisis. In Capital, Marx alternated between two explanations as to why this is. One, perhaps the most readily understandable, is that because machinery enables capitalists to increase production while (other things being equal) reducing their wage bill, it fosters gigantic imbalances between the increasing volumes of commodities produced and the purchasing power available to buy them. This brings on economic stagnation and paralysis, with factories closing and unemployment lines growing, until enough firms go out of business to eliminate the glut of overproduction and get the system moving again – or a social revolution breaks out.

Marx’s other explanation, less intuitively obvious, but arguably more profound, has to do with the tendency to a falling rate of profit (FROP) (1991: 317–38). Because the value of a commodity ultimately depends on the amount of socially necessary labour required for its production, replacing humans with machines lowers the value of the commodity – and hence, eventually, the price it commands, and the profit per item capital can command. Because automation cheapens goods, it makes each single one of these goods less profitable to the capitalist. In the face of this tendency of the rate of profit to fall – a direct result of automation – capital can, at least for a time, maintain the mass of profits by increasing the sheer volume of production, but it is running against its own value-decreasing machinic momentum. Marx (1991: 339–48) listed a number of ways capital can hold this process at bay or even temporarily reverse it, but the tendency of increasingly machinic production is, again, towards spasmodic crisis, this time caused by the flagging profitability of business, and moments of unemployment, immiseration and social tumult.

introduction: ai-capital

17

These two versions of crisis theory, and their degree of compatibility, have been intensely debated. But in both versions, capital’s recurrent crises arise from its inherent drive to substitute constant (machinery) for variable capital (labour). This either reduces consumption power (by cutting wages) or lowers the profitability of production (by cheapening goods), or both. The outcome is repeated throughout deepening cycles of economic breakdown, each of which offers the possibility of a revolution by a working class suffering from downward pressures on wages and the threat of unemployment, but still central enough to production to halt or take control over it. This, then, with its internal bifurcations and attendant controversies, is what we call the major line of Marxist thinking about capitalism and machines. An analysis of AI undertaken within this current will analyse it in terms of labour exploitation, inter-capitalist competition and capitalism’s techno-induced crisis tendencies.

What we dub the two minor tendencies in Marx’s thought are extensions of, but also deviations from, this major line. Latent in Marx’s account of capital’s increasing mechanization is the idea that the positions of worker (initially the main, value creating actor) and the machine (at first the worker’s power-amplifying supplement) invert. The worker, who at the handicraft stage was the subject of the labour process, becomes an automaton of repetitive, repeated motions, responding to automatic machinery rather than using it; the automatic machinery has become the subject of the labour process. Two subsidiary, perhaps maverick, lines of Marx’s thought develop this idea to its logical conclusion – but in two different directions.

The first, and by far most famous, comes from the ‘Fragment on Machines’ in Marx’s Grundrisse, which from the 1970s on has been seen as an extraordinary anticipation of high-technology capital. In the ‘Fragment’, Marx envisaged capital making vast techno-scientific achievements by mobilizing the ‘general intellect’. This enables it to reach a level of automation that, while not eliminating human labour entirely, reduces and relegates it to the peripheral position of supervising a mainly machinic process. This might seem the final triumph of capital over its troublesome working class, but Marx in the ‘Fragment’ presented it as a pyrrhic victory. By removing the necessity to base production on wage-labour (and hence liquidating the possibility of basing consumption on waged income), it undermines value, i.e. the whole basis of capital’s social organization. Automation inadvertently subverts capital by abolishing work. This is consonant with other celebrated passages of Marx’s concerning the liberatory nature of technological development, most notably the account of how developing ‘forces of production’ burst apart fossilized ‘relations of production’, making way for the appearance of a whole new ‘mode of production’ (Marx 1973 [1859]).

18

inhuman power

At first glance, the ‘Fragment’ might seem just to restate in especially emphatic form the predictions about the mounting organic composition of capital that inform Marx’s thinking about the FROP. As George Caffentzis (2013) points out, however, there is a divergence between the two theories, for while the FROP depends for its operation on the validity of the labour theory of value, the ‘Fragment’, in contrast, posits a dynamic of capitalist collapse arising from the liquidation of labour – and value. The Grundrisse was translated into English, French and Italian at the very time when computers were first beginning to enter the workplace, and from the moment of its appearance it was taken as harbinger of a high-technology, ‘cyborg’ communism which would overcome all the difficulties a drab, industrial, actually-existing socialism was experiencing in organizing work on a non-capitalist basis by simply eliminating the need to work at all. It is therefore no surprise that the ‘Fragment on Machines’ is a foundational document for theorists of left accelerationism, postcapitalism and fully automated luxury communism, so much so that Frederick Harry Pitts (2017) has dubbed these all instances of ‘Fragment Thinking’.

A second minor strain in Marx’s thought, less discussed than the ‘Fragment’, is the nightmare vision of capital in the ‘Results of the Immediate Process of Production’ (appearing in some editions of Capital as an Appendix to Volume 1) (1990: 949–1084). In this text, Marx gives an account of the process of capitalist ‘subsumption’ – roughly translated as domination, envelopment or take-over – of labour. He details two moments: formal and real. In the first moment, capital organizes labour as wage-labour, thus merely changing its social form, while leaving the content of labour, i.e. how it is carried out, the same as in pre-capitalist artisanal handicraft. In the second moment, however, the content of labour changes – it is really subsumed – to better meet the dictate and demands of the capitalist production of surplus-value. Initially, real subsumption occurs through introducing a division of labour into the handicrafts, but subsequently by the automation of labour by machinery, which requires capital to absorb socially produced scientific knowledge to develop technologies adequate to and commensurate with its own priorities – notably, the automation of production and the acceleration of the circulation of commodities. In the transition from formal to real subsumption, ‘absolute’ exploitation of labour – the extension of the working day – is displaced by the extraction of ‘relative surplus-value’, which is based on increasing the productivity of labour by intensifying the labour process through the division of labour and machinery. As this process builds, Marx argues that a situation emerges in which, in the industrial factory (only nascently visible in his time), the worker confronts a fully ‘alien power’ that appears endowed with a ‘colossal independence’ from human agency, rendered ‘autonomous’ by techno-science.

introduction: ai-capital

19

This account might seem just another description of the mounting organic composition of capital, or indeed of the semi-automated ‘animated monster’ of capitalist machinery featured in Grundrisse (1993: 470). There are, however, differences in inflection. For one thing, since the subsumption argument is that capital actually adopts machinery that it designs to its systemic requirements (the valorization of value through elimination of human labour and acceleration of commodity circulation), it becomes more difficult to envisage how the ‘forces of production conflict with ‘relations of production’ – if anything, the former would seem to reinforce the latter. There is certainly no hint in the ‘Results’ of either the crisis-inducing falling rate of profit or of the self-destructive labour-abolishing logic of the ‘Fragment’. There is just the towering presence of an all-but-incomprehensible production apparatus that looms over and surpasses the worker it once depended on. This is all the more apparent because, while the first half of the appendix includes one of the most complete discussions in Marx of the powers of the ‘collective worker’ engaged in the cooperative fusion of various types of work, from manual labour to engineering, by the end the collective worker is dwarfed and seems virtually obliterated by the machine apparatus to which its powers have been transferred.

20

inhuman power

It can fairly be argued that, read as an appendix to Capital, this document should be understood only in the light of what precedes it, so that capitalist breakdown can be assumed. And it can further be asserted that the machinic autonomy of capital is only an ‘appearance’ – a mystification that has to be seen through to detect the continuing, if baroquely veiled, importance of labour-power situated down remote production chains. Maybe so. But as even chronic optimists like Antonio Negri (2017) note, ‘appearance’ in Marx doesn’t mean ‘shadowy, superficial or insubstantial’: on the contrary, it means a concrete social reality created on the basis of a mystified and disguised process, namely the incremental subsumption of the power of workers into machinery. While Marx developed this sombre vision through analysis of the industrial factory, it invites thought about a further stage of ‘hyper-subsumption’ in which capital’s autonomizing force manifests as AI: later in this book we explain how this new stage of subsumption is unfolding, and may culminate in AGI.

Both the major and minor lines of Marx’s machine analysis describe dynamics in play in the actually-existing AI-capitalism of 2019. We are not at the heights or depths of machinic capitalism foreseen in either the ‘Fragment’ or the ‘Results’. Workers on Foxconn assembly lines, in Amazon warehouses or on Facebook content moderation sites all attest to the continued use of machines to intensify and speed up human labour. Crises such as the 2008 Wall Street crash reveal the deep instabilities which machinically depressed wages, digitally organized cheap labour and high-speed financial trading are bringing to capitalism. So much of the major line of Marx’s machine thought holds true, perhaps truer than ever, as the 2020s approach. But ML-driven AI, developed in part in response to the crisis of capitalist globalization, is placing on the horizon possibilities that resemble those in Marx’s visions of capitalism’s machinic extremes. The ‘Fragment on Machines’ and the ‘Results of the Immediate Process of Production’ are recto and verso of one page, a page that speaks both of machine power liberating humanity from capital, and of a capital rendered autonomous from humanity. Capitalists and communists alike, be careful what you wish for!

IN THE AGE OF SELF-REPLICATING AUTOMATA

Marx’s discussion of machinery stops with the steam-powered factory. Only towards the end of his life were electricity and electromagnetism harnessed, and although early computer technology like Jacquard’s Loom and Babbage’s failed Difference and Analytical Engines existed when he wrote, he did not discuss them. So to bring Marxism to bear on AI, Marx’s account must be amended. There is a formidable Marxian literature on ‘cybernetic capitalism’ and ‘digital labour’, but Marxian analysis specifically devoted to AI, analysing the specificity of its technology, political economy and class implications, is relatively rare. There are, however, three treatments that have influenced our thinking.

introduction: ai-capital

21