人工智能和资本主义的未来

1. and the Cyc project initiated by Douglas Lenat in 1984 (Lenat and Brown 1984), it was not until the fiftieth anniversary of the Dartmouth workshop and the rise of ML in the early 2000s that the original goal of AI, reincarnated as AGI, has again become acceptable and seen as realistic (Goertzel 2007: 1162).

Despite AGI sharing the original goal of AI, what the former refers to differs significantly from what AI has come to mean as it has been applied commercially in the 2010s. Whereas narrow AI focuses on solving narrowly constrained problems, the goal of AGI research and design is to create systems ‘with sufficiently broad (e.g. human-level) scope and strong generalization capability’ so that it is able to transfer knowledge ‘from one problem or context to others’ (Goertzel 2014: 2–3).

A term that appears to be near-synonymous with AGI is ‘human-level machine intelligence’ (HLMI).³ What is interesting about the definition of this term is that it is grounded in labour.⁴ One definition refers to HLMI as AI that ‘can carry out most human professions at least as well as a typical human’ (Müller and Bostrom 2016: 558), and another posits that it ‘is achieved when unaided machines can accomplish every task better and more cheaply than human workers’ (Grace et al. 2017: 1). A machine that is capable of doing many different types of jobs would almost by definition have to be generally intelligent if it were to master jobs that it had not been taught to do. While HLMI would appear to be a perfect fit to discuss the possibility of increasingly intelligent machines that eventually become capable of labour, we reject this term because of its anthropocentric notion that a machine can be defined as intelligent only if it is human-like. As Benjamin Bratton (2015) argues, ‘human intelligence simply can’t exhaust the possibilities’ of all forms of intelligence because we ‘would do better to presume that in our universe, “thinking” is much more diverse, even alien, than our own particular case’. AGI is therefore a preferable term to HLMI because it ‘stress[es] the “general” nature of the desired capabilities of the system’ (Wang and Goertzel 2007: 1). An AGI could possess intelligence equally general to that of a human being without mimicking human cognition. This difference of cognition is already readily clear with how narrow AI does things compared to humans. For example, ‘AI chess programs will use brute force searches in some instances in which humans use intuition, yet the AI can still perform at or beyond the human level’ (Baum 2017: 8). In addition, AGI is a preferable term because, as Ben Goertzel argues: (1) humans are not necessarily that smart, hence, ‘human-level’ may actually be limiting the scope of what AGI is; and (2) if it is possible to conceive of and thus develop ‘an AGI system with different strengths and weaknesses than humans, but still with the power to solve complex problems across a variety of domains and transfer knowledge flexibly between these domains’, then it becomes tough to define whether the system is human-level (2007: 1163). The arguments made by Bratton and Goertzel are redoubled by those who suggest that what is really at stake in AI is the dissociation of intelligence and consciousness (Shanahan 2015; Harari 2016), that is to say, the possibility that a machine could behave intelligently without possessing a self-awareness similar to that experienced by humans.

112

inhuman power

Let us reiterate: as of 2019, AGI does not exist. Some doubt that such machine intelligence is even possible (Dreyfus 1972; Braga and Logan 2017; Nadin 2018). Within the AI community, however, many consider AGI to be possible at least in theory and some argue that it ‘is merely a very difficult engineering problem’ (Pennachin and Goertzel 2007: 1). At the very least, AGI is considered sufficiently plausible to attract investments for research and business ventures.

Within the larger AI community, AGI researchers form a very small minority. The narrow AI industry dwarfs the AGI one in terms of the number and size of corporations and research institutions, and projects and funding. Baum’s 2017 survey of active AGI projects gives an overview of the political economy of AGI as an industry. This survey identifies a total of 45 active AGI projects⁵ in 30 countries worldwide, with most of them being ‘based in major corporations and academic institutions’ and a smaller number based in public corporations, nonprofits or governmental institutions.⁶ While ‘some ... are large and heavily funded’, most of them are small to medium in scale (Baum 2017: 2). AGI research is characterized by having a large number of open-source projects, with a total of 25 projects having made their source code available online. While only nine of the AGI projects have identifiable military connections, only four were identified as clearly having no military connections with the rest being unclear. Most AGI projects are based in the US or within its sphere of influence, and the only ones that are not are based in Russia and China (Baum 2017: 2).

The largest AGI projects are Alphabet-Google’s DeepMind, the Elon Musk-backed Open AI, and the Human Brain Project, while other notable projects include Vicarious FPC, the Microsoft acquisition Maluuba, Open Cog, Uber AI, and Nnaisense. According to Baum, there are two trends among the for-profit corporations: either to support ‘long-term AGI R&D in a quasi-academic fashion, with limited regard for short-term profit or even any profit at all’, or to undertake ‘AGI projects delivering short-term profits for corporations while working towards long-term AGI goals’ (2017: 19) The latter trend is apparent in corporations like Microsoft, Google and Uber; Baum argues that if ‘this synergy between short-term profits and long-term AGI R&D proves robust, it could fuel an explosion of AGI R&D similar to what is already seen for deep learning’ (2017: 19). Based on Baum’s survey, it appears that the current state of AGI research is still primarily, as Pei Wang and Goertzel argued a decade prior, ‘producing publications and preliminary results’ (2007: 3).

perfect machines, inhuman labour

113

How far off we are from inventing AGI is uncertain, but many working in the field of AI view it as inevitable. Ethem Alpaydin argues that, based on what ML systems can do today, it ‘will not be surprising if this type of learned intelligence reaches the level of human intelligence some time before this century is over’ (2016: xii). There are, however, more optimistic estimates. In a survey asking AI experts when they expected HLMI to be developed, the median estimate was a 10 per cent probability by 2022, 50 per cent probability by 2040, and 90 per cent probability by 2075 (Müller and Bostrom 2016). Another survey asking ML experts when AI will exceed human performance yielded similar results: 10 per cent chance of occurring within nine years, 50 per cent within 45 years, and 75 per cent within 100 years (Grace et al. 2018). This survey also asked granular questions about when AI will outperform humans in a particular skill, giving estimates for many different tasks, including translating languages by 2024, driving a truck by 2027, working retail by 2031, and writing a best-selling book by 2049. The ML experts believed that there is a 30 per cent chance of AI outperforming humans in all of these tasks in 45 years, and that all human jobs would be automated in 120 years.⁷

AGI is actively pursued by corporate, governmental and nongovernmental interests. The actual invention of such machines may be one of the bleakest possible futures of capitalism for humans because, as we now critically discuss, such a machine could labour and therefore potentially also create value. Although being able to perform labour is a necessary condition for creating value, it is not a sufficient one. We therefore proceed analytically and consider the possibility of AGI labouring separately from whether it could create value.

114

inhuman power

HUMAN LABOUR AND LABOUR-POWER

In Capital, Marx presupposed labour as ‘an exclusively human characteristic’ and repeatedly qualified this central concept as ‘human’ and ‘living’ or even ‘living human’ (1990: 283–4). His qualification of labour in this way can be explained as rhetoric against the bourgeois political economists who fetishistically argued that land and capital, in addition to labour, were sources of value.⁸ But Marx’s assumption that labour is inherently human is more than rhetoric. While Marx primarily discussed labour with reference to the social relations in which it occurs, the touchstone for his concept of labour was, for better or for worse, the human being. That labour is human is one of Marx’s axioms.

When Marx discussed labour as such, i.e. without reference to the society in which it occurs, he sharply distinguished it from the productive activity of animals. In Marx’s theoretical framework, when animals are employed in a capitalist production process, unless as raw-material inputs, they are functionally equivalent to machines, reduced to the status of fixed capital.⁹ Given this functional equivalence, the arguments Marx advanced in the $\scriptstyle 1 8 4 4$ Manuscripts and Capital for why labour is qualitatively different in kind from the productive activities of animals are salient for understanding both why machines cannot labour – and why, in theory, AGI could labour. Marx’s arguments for why animals cannot labour are especially interesting because they describe animals in terms quite similar to the domain-specific behaviours of narrow AI. By extension, this suggests an isomorphism between the concept of AGI and Marx’s concepts of labour and labour-power.

In the seventh chapter of Capital, Marx argues in transhistorical terms that labour is ‘a process between man [sic!] and nature, a process by which man through his own actions, mediates, regulates and controls the metabolism between himself and nature’, and through this metabolism the human species ‘appropriate[s] the materials of nature in a form adapted to [its] own needs’ and ‘acts upon external nature and changes it’ $\left( 1 9 9 0 \colon 2 8 3 \right)$ . The statement makes it appear as if Marx is arguing that humans are the only beings that act on, use and alter their lived environment. In the $\scriptstyle 1 8 4 4$ Manuscripts, however, he recognizes that all animals are ‘sensuously acting, objectifying subjects’ that build their own versions of the world and reproduce their habitats, albeit in a speciesspecific manner (Fracchia 2017). Throughout the $\scriptstyle 1 8 4 4$ Manuscripts, Marx recognizes that humans and animals have a lot in common because all living beings must engage in a ‘continuous interchange with nature’ in order to survive (Mulhall 1998: 10). In addition to producing speciesspecific habitats, humans share with animals the ‘animal functions’ of eating, drinking and procreating, etc. (Marx 2007: 75). But despite the similarities, the specific way in which humans engage with their external environment differs significantly from that of animals.

perfect machines, inhuman labour

115

Marx explains this difference with the concept of life-activity and argues that the ‘whole character of a species … is contained in the character of its life-activity’. Whereas an animal ‘is immediately identical with its life-activity … does not distinguish itself from it [and] is its life-activity’, humans are distinguished by having ‘conscious life-activity’ (2007: 75). One of the dividing lines Marx thus draws between humans and animals is that the former are conscious of what they do, while the latter are not. He reiterates this argument in Capital, stating that the conscious purpose behind what humans do ‘determines the mode of [their] activity’ (1990: 284). In turn, this consciousness gives humans ‘free play of [their] own physical and mental powers’ (1990: 284); they are therefore ‘free’ in their activity (2007: 75).

Why does consciousness grant human labour-power freedom? What does Marx mean by ‘free’? The implication is, of course, that in their activity animals are unfree. But why? This difference in relative freedom can be better explained by turning to two passages where Marx underscores the difference between humans and animals. Importantly, these passages also identify other mental capacities Marx argued were uniquely human. In the 1844 Manuscripts, Marx writes:

Admittedly animals also produce. They build themselves nests, dwellings, like the bees, beavers, ants, etc. But an animal only produces what it immediately needs for itself and its young. It produces one-sidedly whilst man produces universally. It produces under the dominion of immediate physical need, while man produces even when he is free from physical need and only truly produces in freedom therefrom. An animal’s product belongs immediately to its physical body, while man freely confronts his product. An animal forms things in accordance with the standard and need of the species to which it belongs, whilst man knows how to produce in accordance with the standard of every species and knows how to apply everywhere the inherent standard to the object. Man therefore also forms things in accordance with the laws of beauty. (Marx 2007: 75–6)

116

inhuman power

This should be read alongside the (in)famous bees and architect passage from Capital:

A spider conducts operations which resemble those of the weaver and a bee would put many a human architect to shame by construction of its honeycomb cells. But what distinguishes the worst architect from the best of bees is that the architect builds the cell in his mind [Kopf] before he constructs it in wax. At the end of every labour process, a result emerges which has already been conceived by the worker at the beginning, hence already existed ideally [ideell]. (Marx 1990: 284)¹⁰

Here, Marx homes in on consciousness and cognition (e.g. creativity, learning and imagination) as uniquely human and opposes these to the instincts of animals, which determine their entire life-activity. Both human and animal production is hardwired to the survival instinct, but instinct also determines what and how, and the times and places animals produce and consume. The form and content of an animal’s existence is thus determined by what species it is because all animals are identical with their life-activity (Mulhall 1998: 12; Fracchia 2017). For Marx, animal activity is thus ‘unfree’ because it is instinctual.

Humans thus differ from both animals and algorithms because they have ‘cast off’ the ‘instinctive forms’ of their life-activity, i.e. labour (Marx 1990: 283). The anthropocentrism of Marx’s concept of labour can be understood as based on an assumption Marx made about the evolution of humans and of other animals: the former evolved a consciousness and a series of other mental capabilities while the latter did not. Therefore, humans can make their life-activity and labour ‘an object of will and consciousness’, meaning that they can reason about what, how, when and why they produce, and can imagine and deliberate on what they could do differently (Marx 2007: 75; Mulhall 1998: 13). It is on the basis of such choices humans can make in production that Marx makes the argument that humans have ‘free play’ of their mental and physical capabilities.

This free play also enables humans to produce ‘universally’, i.e. across multiple domains, as opposed to the one-sided, domain-specific production of animals. In other words, humans have a broader range of productive activities than other animals, a range that includes producing ‘in accordance with the standard of every species’. In addition, human production is not limited to immediate physical need because whatever natural limits were once imposed by instinct are overcome and ‘artificially expanded’ (Fracchia 2017). In principle humans can, according to Marx, put their minds and hands to almost anything, including imitating nature and learning from what other animals do and make, such as bees honeycomb cells, beavers’ dams and spiders’ webs, and ‘produce as artifacts that which “immediately belongs to the physical body” of an animal (clothes as artificial skin) and can learn and make things to do what other species do “instinctually” (winged vehicles to fly)’ (Fracchia 2017).

perfect machines, inhuman labour

117

But why can humans do all of this with their labour? Marx’s argument that humans are able to produce universally should be understood in terms of them being able to, first, learn from what they have observed in nature about what animals do, and from what they have succeeded or failed at in the past, and then, second, through their creativity or imagination, apply this new knowledge to a different domain of production, or even invent a new one. In turn, this ability is dependent on the various mental capabilities or phenomena that Marx understands as uniquely human. In the two passages quoted above, Marx explicitly or implicitly grounds human labour in: aesthetic appreciation; learning and understanding (from what other animals produce); a conscious mind; forming ideas; imagining and conceiving plans; and, in general, being creative and adaptive when solving the problem(s) of production. In the aggregate, these mental capabilities can be understood as general intelligence, which makes humans flexible in what they can do and adaptable to new and different environments, objects of labour, and other conditions of production.

HUMANS, ANIMALS AND MACHINES TODAY

There are many problems with Marx’s assumptions and arguments about both humans and animals, as well as the work of architects. At best they were based on nineteenth-century evidence from the nascent science of biology, at worst on his imagination and humanist beliefs. Can the sharp distinctions he draws between humans and animals be sustained in the light of contemporary ethology? Why cannot animals labour? Can Marx’s assertion that labour is inherently human be sustained? The debate about whether animals can labour and/or produce value is decades-old (see e.g. Benton 1988, 1993, 2003; Elster 1985; Wilde 2000; Perlo 2002; Drake 2015; Johnson 2017; Kallis and Swyngedouw 2018; Morton 2017), and Marx’s sharp distinction between the freedom of human labour and the unfree instinctual-algorithmic behaviour of animals has been critiqued using evidence from ethology, i.e. the study of animal behaviour and cognition. Jon Elster (1985) and Ted Benton (1988, 1993), for example, argue that animal production is far from one-sided, being characterized not only by a high degree of variability but also by considerable adaptability, especially when they are confronted with new environmental conditions (such as urban settings). Similarly, Paul Cockshott and Karen Renaud challenge the sharp distinction Marx draws between humans and animals on the basis of planning or having a vision (i.e. imagining) before they do something, noting how wolf packs divide tasks while hunting, how various mammals use episodic memory to plan, and how experiments have shown that spiders form plans while web-weaving (2016: 21–2).

118

inhuman power

Animal behaviour is more general than Marx believed and so is their cognition. Indeed, researchers of cognitive evolution no longer posit that animal behaviour can be fitted into a hierarchy of complexity, but instead focus on identifying and describing animals’ ‘mental organs’ or modules that each have a domain-specific function and evolved to only accept a ‘certain kind of input’ (e.g. optical, aural or olfactory) in relation to a given ecological problem (Vitti 2013: 394). While initially specialized, ‘modular cognition’, due to increasing encephalization, evolved into ‘flexible and domain-general’ cognition (Vitti 2013: 395). What has driven this evolution may be ‘omnivorous diets, variable habitat, long developmental periods, extended longevity, and pervasive social interactivity’ (Vitti 2013: 396). While most animals, the honeybee included, ‘specialize in the skills that their ecological niche requires’, some species, with high encephalization levels (mammals, birds and cephalopods), are considered ‘generalists’ (Vitti 2013: 395).

Timothy Morton (2017: 56) takes a more fundamental issue with Marx’s argument that the labour of human architects is different from that of bees. He contends that it is impossible to demonstrate that either humans or bees have the capacity of imagination, or that humans are not executing algorithms when they plan. The assumptions Marx makes to argue that human productive activity is uniquely distinct from that of animals is, Morton argues, ‘based on a metaphysical assertion about humans’ (2017: 56).

What these critiques make clear is that the distinction between humans and animals is more blurred than Marx imagined. We argue that the distinctions between humans and intelligent machines are also becoming increasingly blurred. Indeed, many of the examples of the complex behaviour and inner lives of mammals and insects that are presented against Marx’s insistence that labour is inherently human have analogues in the functioning of ML systems. If Marx’s conceptualization of labour as human can be challenged with examples of animals doing things previously thought uniquely human, the same holds for machines, especially those that exhibit behaviours that appear to be productive, creative, imaginative and even adaptive. Some ML systems even appear to envision plans or concepts in their neural architectures that are then respectively carried out or recognized in new corresponding examples the system has never before encountered.

perfect machines, inhuman labour

119

A striking example of machine creativity is DeepMind’s AlphaGo, which first beat European Go champion Fan Hui in October 2015 and then Go world champion Lee Sedol 4–1 in a match in March 2016. The ancient board game of Go has long been a milestone in AI research due to it being one of the most complex games humans have designed. Chessplaying AIs can simply use brute force and search all potential moves until the winning ones are found, but this is not possible in Go since there are more possible configurations of the board, and thus potential moves, than there are atoms in the universe. Until AlphaGo proved otherwise, the game was seen as requiring human-level pattern recognition and something akin to intuition.¹¹ To learn how to play Go, AlphaGo relied on a combination of deep neural networks and Monte Carlo tree search. Its architecture consists of three convolutional neural networks (CNN) that were divided into one ‘value network’ and two ‘policy networks’.¹² The CNNs were trained on image data that represented individual moves (‘states’) from 160,000 games recorded from top Go players (Silver et al. 2016). The value network learned how to play the game and to estimate the probability of winning given any current state of the game, while the policy networks learned what move to select given the current state of the game. In turn, the two policy networks’ decision-making was reinforced by having them play each other 30 million times and keeping track of the best outcome of the games (Silver et al. 2016).

AlphaGo’s creativity or intuition was evident in the 37th move of the second game against Sedol. Although this move was later described as ‘creative’, ‘unique’ and ‘beautiful’, as well as ‘inhuman’ by professional and champion Go players (including Sedol), it at first took everyone by surprise (Metz 2016). After AlphaGo made this move, Sedol took an unusually long time to respond. Commentators noted how strange it was, some thinking that the neural network had made a mistake because a human would never make such a move. But the move was deliberate and based on what AlphaGo’s policy networks had learned. The engineers at DeepMind could not have predicted this move and could only determine AlphaGo’s reasoning after it had made its move: having calculated the probability of a human making this particular move at that particular state of the game at 0.010 per cent, it nevertheless made it (Metz 2016). AlphaGo arguably made a plan and executed it. Not only did AlphaGo learn how humans play, but it went beyond the constraints that generations of human Go players had placed on how to play the game, making seemingly counterproductive and highly novel moves, and playing the game in styles that no human would (Metz 2016). AlphaGo’s creativity comes from moves like this and from how it learned to play Go in a completely new and inhuman style; if a human had done any of this, she would likely have been recognized as original, creative and highly intelligent.

120

inhuman power

Despite this victory, we note that ‘Go’ in Katja Grace et al.’s (2018) list of AGI milestones has not been passed. This milestone is specifically listed as ‘Go (same training as human)’ (Grace et al. 2018: 747). While these shifting goalposts can be explained in terms of the AI effect, AlphaGo’s victory, while impressive, is less so when compared to how a human learns how to play the game. As Brenden M. Lake et al. (2017) point out, a human being can learn Go after playing just a few times, but for a ML system to do the same, it requires data from hundreds of thousands to millions of games. AlphaGo was trained on over 30 million games of Go, whereas Lee Sedol, has likely played only 50,000 games in his lifetime (Lake et al. 2017: 23). This difference emphasizes the major drawback of machine learning: the excessive amount of data it needs to do just one thing. On the other hand, human beings are able to ‘build rich causal models, make strong generalizations, and construct powerful abstractions’, all based on extremely limited data (Tenenbaum et al. 2011: 1279). Indeed, the capacity to generalize based on just a few examples may be at the core of general intelligence.

Some cognitive psychologists suggest that human beings generalize from limited data by using their imagination; a path towards AGI may thus lie in overcoming the problem of data gluttony by equipping systems with artificial imagination (Lapovsky 2014; Johnston 2008: 408; Ritter et al. 2017). Imagination is yet another factor that Marx argued marked labour as uniquely human: the human architect builds something in the mind before construction begins in the real world.

perfect machines, inhuman labour

121