人工智能和资本主义的未来

The start-up Vicarious PFC has built a system for solving the CAPTCHA codes (‘Completely Automated Public Turing Test to tell Computers and Humans Apart’) that are specifically designed to distinguish humans from online robots. It rests on giving AI what Vicarious claims is an ‘artificial imagination’ (George et al. 2017). According to this company, imagination concerns the ‘ability to picture what the information [the AI has] learned should look like in different scenarios (Knight 2016a). More precisely, the company ties such an imagination to the inductive biases and constraints in the visual cortex and neocortex, which cognitive psychology and neuroscience have shown are connected to the learning efficiency and generalization ability of the human brain (Tenenbaum et al. 2011; George et al. 2017: 1).¹³

Inductive biases – also referred to as learning biases – have been defined as ‘factors that lead a learner to favour one hypothesis over another that are independent of the observed data. When two hypotheses fit the data equally well, inductive biases are the only basis for deciding between them’ (Griffiths et al. 2010: 1). In other words, inductive biases are heuristic processes which the human brain has evolved to deal with the problems of perceiving with and learning from limited data and extrapolating learned experiences to novel contexts. Cognitive and developmental psychologists have proposed a number of inductive biases that human children exhibit, some of which have been emulated in AI projects to identify words and objects (George et al. 2017; Ritter et al. 2017). One inductive bias is the recognition of shapes, which humans prefer over colour, texture or size to categorize objects and words. A predisposition to recognize shapes, and thus identify useful objects, has obvious evolutionary benefits. Shape bias can be understood as an innate interpretive tool that children use when learning new words.

It was through simulating such inductive biases that Vicarious equipped their CAPTCHA-solving AI with an artificial imagination. CAPTCHA codes consist of letters crowded together in distorted combinations and often have added clutter in order to confuse ML algorithms, such as classifiers. These algorithms work well if individual characters can be segmented, but doing this ‘requires an understanding of the characters, each of which might be rendered in a combinatorial number of ways’ (George et al. 2017: 1). Achieving such an understanding to solve even just one specific style of CAPTCHA typically requires a system to be trained on millions of labelled examples of letters in various combinations of shape and orientation. If not, the system has difficulty in telling what the letter is, where it ends, and where the next one starts. Humans, of course, can solve a new style of CAPTCHA without any explicit training because of the inductive biases of the visual cortex. The feedback connections in the visual cortex of mammalian brains enable mammals to identify a figure from the background (i.e. figure-ground perception) and to isolate ‘the contours of an object even when partially transparent objects occupy the same spatial locations’ (George et al. 2017: 1).

122

inhuman power

What Vicarious has referred to (or slyly marketed) as imagination is more generally referred to in the AI community as ‘one-shot learning’, which is a human capability enabled, at least in part, by inductive biases. One-shot learning refers to a human child or ML system being shown a single labelled example and from this being able to learn how to identify other examples of the same category of objects, be it cats, letters or whole words, by relying on various innate inductive biases. For example, the typical human two year old can learn and start to use a new word like ‘cat’, and, after seeing just one example, can recognize a cat from various angles, cats of different colours and sizes and in various poses, and so on. For a ML system to recognize cats in photos, it has to be shown millions of examples of cats from different angles, of different shapes, sizes, materials, colours, and so on (Tenenbaum et al. 2011: 1279; Lake et al. 2017: 22). Similarly, after playing just one or a couple of games of Go, almost any human would have learned how to play the game and could also generalize this knowledge to play one of the many variants of Go with little difficulty, but for AlphaGo to do the same would ‘require significant reprogramming and retraining’ (Lake et al. 2017: 23). Researchers from DeepMind have found that the one-shot learning models of matching neural networks trained on ImageNet also exhibit biases towards shape, although the magnitude of the bias depends on the particular neural architecture (Ritter et al. 2017).¹⁴ Through this inductive bias the neural network learns how to learn categorization or how to be a classifier, i.e. an algorithm that implements classification.

Vicarious simulated inductive biases with a ‘structured probabilistic generative model framework’ on what they call a Recursive Cortical Network (RCN). Generative models produce representations or abstractions from observed or learned phenomena based on a probabilistic model. To explain using Marx’s architect, if she wanted to build a honeycomb-like structure, this architect would have observed honeycombs in a bee’s nest; the memory of these honeycombs would be the architect’s generative model in her mind, from which she can generate (i.e. imagine) many different honeycomb architectural drawings. In AI, such generative models are first created based on the patterns it has recognized in the data and are then used to predict probabilities from this model, such as producing new variations of the data it has been trained on. In the case of Vicarious’ RNC, its inductive bias enabled it to learn the contours and surfaces of alphabetic letters based on little training data and then, in turn, break CAPTCHAs by generatively segmenting characters (George et al. 2017). This generation can be understood as an artificial imagination because segmenting letters in this way is akin to imagining what they would look like in different shapes, orientations and positions or without a particular background or distortion. In this case, imagination appears to be connected not only to an act of creation, but also to adapting to different situations or noisy contexts. For Vicarious, the end goal of equipping AI with an imaginative capacity is to enable it to transfer learning to a completely different domain. Importantly, this transfer can be understood as adaptability, and this word is what AI researchers use to explain how AI can become generally intelligent and/ or more human-like (Alpaydin 2016: ix, xii, 17, 24; Kaplan 2016: 49–50).

perfect machines, inhuman labour

123

WHAT LABOURS?

The differences between humans and animals are not as sharp as Marx believed, and the distinctions between human and machines are becoming blurrier the more sophisticated ML systems become. However, do any of these examples of animal and machine production or seemingly imaginative, creative and adaptable behaviour actually challenge Marx’s assumption that labour is an exclusively human characteristic? Proposing that Marx’s anthropological conceptualization of labour and value creation is void by pointing out that whatever he refers to as uniquely human is also found in the wider animal kingdom is, as Lawrence Wilde argues, no more than a rhetorical strategy whereby a scientific source is used to ‘blur the distinction in productive techniques without genuinely getting to grips with Marx’s argument’ (2000: 44). It would also be a mistake to equate the narrow capabilities and particular functioning of even the most sophisticated deep-learning system with labour. While Marx’s anthropocentrism must be challenged, merely listing a set of behaviours or productive activities does not overturn his argument. As long as AI systems can only perform one or even a few narrowly defined tasks, it does not matter how creative or imaginative they appear to be, their behaviour is not labour.

124

inhuman power

The reason Marx gives for why human labour is different from the activity of both animals and machines is that it is highly flexible, adaptable – in short, general. And while many animals are more adaptable than Marx thought, and have domain-general intelligence, human intelligence is yet more general, which is what gives this particular species a wider qualitative range of productive behaviours than any other. Although Cockshott and Renaud (2016) seek to blur distinctions between what humans and animals can do, they nevertheless recognize that the qualitative range of what humans can do is wider than those of all other animals and current machines. Noting how easily Watt’s steam engine replaced horses as a source of power in production, and that humans are not as easily replaceable (although they recognize that machines do take their place), they argue that ‘there is more to work than muscular energy’ and, hence, that human beings ‘can be set to almost any labouring task … they are adaptable … it is this adaptability which must be a fundamental reason why we are the dominant species. We are, in effect, universal robots’ (Cockshott and Renaud 2016: 21). Thus, when Marx argues that humans have a capacity to labour, he is implying a qualitative range of productive and adaptable behaviours, which includes potentially new ones.

Machines, animals and humans can be placed on a continuum of intelligence ranging from general to narrow, with humans located close to generality, while existing machines, including ML systems, as well as species of insects, lie towards the opposite end. Marx’s concepts of labour and labour-power can also be located somewhere along this continuum: for a being to have the capacity to perform labour, it must reach some as yet undefined critical mass of encephalization and threshold of generality of intelligence that, as of 2019, only human beings have so far crossed. Thus, while we agree with Marx that it is still only human beings that can labour, by placing the capacity to labour on a continuum of degree of generality of intelligence, we nevertheless argue that another being could labour and, therefore, that Marx’s concept of labour should no longer be considered necessarily anthropological. Grounding the concept of labour in general intelligence severs its inherent connection to human beings and, therefore, theoretically allows for the possibility that some other being that is generally intelligent could labour. The actual emergence of such a being would empirically prove that labour is not an exclusively human characteristic.

perfect machines, inhuman labour

125

Given that AGI researchers understand general intelligence to refer to the ability to reason with general knowledge, to perform many different tasks, and to operate in new and completely different domains and environments, Marx’s arguments for why only humans can labour can be reformulated as: human beings can labour because they are generally intelligent. We therefore posit an isomorphism between general intelligence (at least how the AGI community understands it) and Marx’s concepts of labour-power and labour; if general intelligence is isomorphic with Marx’s concept of labour-power, it follows that AGI, almost per definition, would be capable of performing labour.

A possible objection against the possibility of AGI labour is based on the claim that machines, following John Searle’s (1980) Chinese Room argument, can never be conscious. Given that Marx considered labour to be conscious and purposeful activity, Searle’s argument is also an objection against the possibility of AGI being capable of labouring. While it may be possible to automate the mental capabilities of creativity and imagination and a degree of adaptability, consciousness may prove to be the last-ditch defence against machines further encroaching on human uniqueness. But to what degree does consciousness or self-awareness matter?

For some, it is not really possible to talk about intelligence unless the machine is conscious and can, therefore, understand what it is doing in the sense of its actions and thoughts being meaningful (see e.g. Searle 1980; Nadin 2018). In relation to creativity, for example, it has been posited that an intelligent machine should be able to consciously identify what is actually novel or valuable (be it cultural or economic) about its creations or its creative process, and be able to explain the difference between bad and good art (Coeckelbergh 2017). Others posit that an embodied AI that has to navigate not just a particular environment, but the world in general, would likely require some sense of self (which is tied to consciousness). For example, Christoph Adami argues that embodied AI ‘have to ascertain where they are in the world, and like us, they would work better, if they have an accurate sense of self’ (quoted in Johnston 2008: 411).

But to what degree is intelligence really tied to consciousness? In Homo Deus, Yuval Noah Harari suggests that ‘intelligence is decoupling from consciousness’ (2016: 101). He notes that until quite recently there were certain tasks – playing chess, driving cars or diagnosing diseases – that not only required a lot of intelligence, but could only be done by conscious human beings. IBM’s Deep Blue beat the then-reigning world champion Garry Kasparov in a six-game chess match in 1997 through a brute force calculation of possible moves, and while diagnosing diseases and driving cars require more than brute force, machines need not be conscious as long as their neural architecture is deep enough to recognize fine-grained patterns in data. Similarly, the AI researcher Murray Shanahan (2015) suggests that ML may be on a trajectory that divorces intelligence from self-aware consciousness. Following Harari and Shanahan, we argue that labour can also be decoupled from consciousness. Whether the human being should be the touchstone for creativity, imagination and intelligence – even consciousness – is beside the point; all that would matter to capital is that these capabilities are emulated. In the age of intelligent machines, the Chinese Room may very well be the hidden abode of production (Kjøsen 2018: 161).

126

inhuman power

TOWARDS AGI

Ethem Alpaydin states that he would not be surprised if a system ‘reaches the level of human intelligence’ through machine learning (2016: xii). Such a view is shared by, among others, Andrew Ng, who advocates the view that ‘human intelligence boils down to a single algorithm’ and believes that deep learning could, if not solve the problem of emulating human intelligence, at least come closer to achieving this goal than anything that has been tried before (Domingos 2015: 117). Similarly, Domingos identifies five main algorithmic ‘tribes’ in machine learning, but asserts that the goal is to unify ‘the key features of all of them’ into ‘the ultimate master algorithm’ (2015: xvii), and declares that AI exceeding human intelligence will be reached soon after this goal is reached (2015: 286).

Such narrow AI approaches to general intelligence can be understood as viewing intelligence as a ‘toolbox’ containing mostly unconnected tools. But connecting them may be difficult due to each tool having been built ‘according to very different theoretical considerations’, thus to ‘implement them as modules in a big system will not necessarily make them work together, correctly and efficiently’ (Wang and Goertzel 2007: 6). While recognizing the possibility of an integrative approach, the AGI community disputes the possibility of general intelligence emerging out of narrow approaches or that AGI could even bridge the gap between various narrow AIs (Wang and Goertzel 2007: 6; Goertzel 2014). As Goertzel argues, the ‘AGI approach takes “general intelligence” as a fundamentally distinct property from task or problem specific capability, and focuses directly on understanding this property and creating systems that display [it]’ (Goertzel 2014: 2). Recognizing that both the narrow AI and AGI fields of research agree that learning is a key aspect of intelligence, Wang and Goertzel point out that

perfect machines, inhuman labour

127

most of the existing ‘machine learning’ works do not belong to AGI … because they define the learning problem in isolation, without treating it as part of a larger picture. They are not concerned with creating a system possessing broad-scope intelligence with generality at least roughly equal to that of the human mind/brain; they are concerned with learning in much narrower contexts. (Wang and Goertzel 2007: 2)

Similarly, Nils J. Nilsson argues that reaching the goal of HLAI through ‘building special-purpose systems’ may be misplaced and instead advocates a ‘general-purpose educable system that can learn and be taught’ and that has ‘minimal, although extensive, built-in capabilities’ (2005: 68). These capabilities include a sensory-motor system, predicting and planning, learning, and reasoning and representation. In effect, Nilsson argues for a ‘child machine’ that learns and develops in a similar fashion to humans.

Nevertheless, while the pursuit of AGI is different from narrow AI, it is not fundamentally so. An AGI researcher may make use of methods from machine learning that concern generalization (e.g. transfer learning and inductive biases) and thus overlap with the goal of creating a general intelligence, but ‘additional architectural and dynamical principles would be required, beyond those needed to aid in the human-mediated, machine learning aided creation of a variety of narrowly specialized AI … systems’ (Goertzel 2014: 4). At the same time, however, a ‘general purpose’ computer system does not necessarily require a ‘single factor … to be responsible for all its cross-domain intelligence. It is possible for the system to be an integration of several techniques’ (Wang and Goertzel 2007: 5). That is, although Wang and Goertzel reject the possibility of aggregating various narrow AIs into an AGI – which would be the equivalent of joining many different automated concrete labours into labour-power – they appear to be open to the possibility that, as long as the general nature of the system is kept in mind while solving narrower problems, they could then be integrated into a general intelligence. Whether a machine with general intelligence can emerge from the current and future refinement of the deep learning approach, or whether the system must be built for a general purpose from the get go, is an ongoing debate in the AI community. It is far beyond our expertise to judge one as more probable or promising than the other.

128

inhuman power

But how seriously should we consider the possible emergence of AGI given the nascence of the field of AI? It is possible that, as Kurzweil (2005a: 33–4) and Bostrom (2014: 29) have suggested, AGI could emerge through a process of recursive self-improvement whereby an AI with access to its own design, and an ability to upgrade it, makes an improved or entirely new version of itself (e.g. with a completely different neural architecture or moving from parallel to quantum computing), which in turn improves itself again ad infinitum. This process would also accelerate, meaning that improvements in machine intelligence would come at closer and closer intervals, and since recursive self-improvement would also be done by AGIs, it could lead to the emergence of artificial superintelligence (ASI).

This argument is not entirely foreign to Marx’s thought. As we discussed in Chapter 1, Marx argued that large-scale industry did not have ‘an adequate technical foundation’ and could not ‘stand on its own feet’ until the general conditions of production had changed so that almost everything, including machines, was produced by machines (1990: 506). Marx understood this phenomenon to be the central characteristic of the industrial revolution. It is possible that, for AGI to emerge, capitalism must go through another revolution akin to the original industrial revolution in order to create a technical foundation that is adequate to ‘fully developed AI-capitalism’. AI must become a general condition of production such that AI is not just produced by means of AI, but that AI also recursively improves itself.

Google’s AutoML (automated machine learning) project is an example of such a process because it is a ‘machine-learning algorithm that learns to build other machine-learning algorithms’ (Metz 2017b).¹⁵ When researchers build a neural network they have to run hundreds of experiments on powerful computers to test which of countless possible permutations works best; this process requires adjusting the model and its parameters many times over, and researchers cannot always precisely explain why they make this adjustment over that; an undefined sense of intuition is often the only guide. It is this difficult process that Google aims at automating by creating ‘algorithms that analyse the development of other algorithms, learning which methods are successful and which are not’ (Metz 2017b). While a far cry from recursive self-improvement, it is part of a significant current trend in AI research that is also important to AGI, namely the focus on how algorithms can learn to learn.

perfect machines, inhuman labour

129

AGI is a completely different type of machine both from the nineteenth-century steam-powered machines Marx discussed in Capital and from the narrow AI of the early twenty-first century. Being generally intelligent, it would have a generic capacity for labour and would therefore be capable of performing labour. While labouring machinery and value-positing machinery are Marxist oxymorons, Marx seems to recognize this possibility in a little-known and curious passage from the Grundrisse that appears about 60 pages after the ‘Fragment on Machines’.

THE FRAGMENT ON PERFECT MACHINES

Did Marx have a theory of AI or intelligent robots? This may seem a strange question to even entertain given that nothing even remotely resembling those technologies existed during his lifetime. However, the notion of creating artificial humans and animals long pre-dates the birth of Marx. The various automata of the eighteenth century were celebrated and widely known. Jacques de Vaucanson built the automata ‘The Flute Player’ and ‘Digesting Duck’ in the mid-eighteenth century, Pierre Jaquet-Droz built his three doll automata between 1768 and 1774, and during the Victorian era people were so fascinated with these mechanical devices that the period from 1848 to 1914 has been described as the ‘golden age of automata’ (Bailly 2003). It is therefore likely that Marx was aware of such automata and they may have been in the back of his mind when he wrote the following:

If machinery lasted for ever, if it did not itself consist of transitory material which must be reproduced (quite apart from the invention of more perfect machines which would rob it of the character of being a machine), if it were a perpetuum mobile, then it would most completely correspond to its concept. Its value would not need to be replaced because it would continue to last in an indestructible materiality … It would continue to act as a productive power of labour and at the same time be money in the third sense, constant value for-itself. (1993: 766)

130

inhuman power

On the surface this passage is unremarkable in the theoretical point it makes about the durability and cost of maintaining machines: an indestructible machine could never completely transfer its value into circulation, which effectively means that because its value would approach zero, it could continuously function to reduce necessary labour time and produce relative surplus-value without any additional outlay of capital.¹⁶ But what makes this passage even more interesting is that Marx’s invocation of a perpetuum mobile places it in the realm of science fiction. No such machine, of course, existed during Marx’s time, nor will one ever exist according to the first and second laws of thermodynamics. Thus this passage can be interpreted as Marx engaging in a flight of fancy as he did at several other points in Grundrisse, including the ‘Fragment on Machines’. If Marx was in a speculative mindset, what he states in the bracketed part of the quote becomes more interesting than his mention of perpetual motion: ‘quite apart from the invention of more perfect machines which would rob it of the character of being a machine’. As one of us (Kjøsen 2013a, 2013b, 2018) has argued through a science-fictional optic, this statement is the closest Marx comes in his oeuvre to thinking about the possibility of something like intelligent robots, androids or AGI. Indeed, Marx could have been speculating about androids, which, while a rare word, already in 1837 meant an ‘automaton resembling a human being in form and movement’ (Online Etymology Dictionary n.d.) and was often used in reference to automated chess players. Marx’s reference to ‘perfect machines’ can, therefore, be interpreted in the following way:

Given the context of Marx discussing fixed capital and knowing that no machinery can create value, if a machine’s character of being a machine is robbed of it, it means that it negates its own being as fixed capital and becomes its opposite, namely variable capital. The perfect machine is dead labour resurrected as living labour … The perfect machine is a machine that can create value, but for that reason it is no longer a machine. (Kjøsen 2018: 173)

It is, however, necessary to unpack Marx’s argument in more depth because although it is theoretically possible that a machine with a general intelligence could perform labour and be living labour, it does not automatically follow that it could create value.

perfect machines, inhuman labour

131