人工智能和资本主义的未来

INFRASTRUCTURAL AI

Our approach to the question of how AI might become part of the general conditions of production is informed by the ‘infrastructural turn’ in the humanities and political economy (Rossiter 2016; Cowen 2014; Steinhoff 2019a), as well as Marxist assessments of logistical and energy infrastructure (Toscano 2011; 2014; Bernes 2013; Kjøsen 2016). Such critical approaches seek to counteract the frequent invisibility of infrastructure by showing how it is implicated in varieties of power relations. However, the particular notion of infrastructural AI comes to us directly from the representatives of AI-capital. As we have seen, commentators such as Andrew Ng and Kevin Kelly expect AI to become ubiquitous, distributed by a network infrastructure, just as electricity and internet access are distributed today. Is there a fetter to production or circulation, a supply bottleneck or something else that could be motivating this particular situating of AI? No less than Microsoft (n.d.) identifies one for us in what it terms the ‘fundamental constraint’ of human cognitive limitations. Framing modernity as an information explosion which escalates with the computer, Microsoft laments: ‘In the midst of this abundance of information, we’re still constrained by our human capacity to absorb it.’ The notion is that the shift to a data-centric mode of production is underway and that, as Marx argued about large-scale industry, the tech industry will not be able to ‘stand on its own feet’ until it creates for itself an ‘adequate technical foundation’ (1990: 506). This foundation is infrastructural AI – the means of cognition.

The slogan under which major AI producers advance their creation of a generalized AI infrastructure for advanced capital is ‘the democratization of AI’ (Microsoft n.d.; Gosaduaff 2017; Gent 2018). Microsoft, Google and Amazon have all announced projects with this goal. Microsoft has announced a plan to ‘democratize Artificial Intelligence (AI), to take it from the ivory towers and make it accessible for all’. It lists the following four points, which deserve to be quoted in their entirety:

We’re going to harness artificial intelligence to fundamentally change how we interact with the ambient computing, the agents, in our lives.

52

inhuman power

We’re going to infuse every application that we interact with, on any device, at any point in time, with intelligence.

We’ll make these same intelligent capabilities that are infused in our own apps – the cognitive capabilities – available to every application developer in the world.

We’re building the world’s most powerful AI supercomputer and making it available to anyone, via the cloud, to enable all to harness its power and tackle AI challenges, large and small. (Microsoft n.d.)

Microsoft imagines a world submerged in AI that is nothing short of techno-animistic: ‘As we infuse intelligence into everything, whether it’s your keyboard, your camera, or business applications, we are essentially teaching applications to see, to hear, to predict, to learn and take action.’ Other tech giants have used the same terminology. Guy Ernest from Amazon describes Amazon Web Services as ‘democratizing AI’ by making their AI tools available for ‘any team size and skill, and for every use case’. The very same phrase has been uttered by Fei Fei Li, Google’s Chief Scientist of the Cloud and ML, and Michael Marin, a senior executive at IBM Internet of Things (Greene 2018; Simpson 2018). ‘Democratizing’ AI thus means generalizing both its deployment and the tools for creating it, making it increasingly available to end-users and allowing anyone, working in any field, even those without any AI training, to develop AI.

One of the most significant axes of the ‘democratization’ programme might seem to contradict the AI industry’s profit orientation. The AI industry is characterized by a large and vigorous open-source community, in which tools and templates for making AI are freely distributed, projects are undertaken by cooperative online programming collectives, and products are released gratis for general use. Nearly all of the tech giants have open-sourced some of their AI-related materials (Simonite 2015; Crosby 2018). In 2015, Google released TensorFlow, a library of tools for deep learning programming, under the Apache 2.0 open-source licence, and it is now widely used. Other tech giants, seeing Google’s success, followed suit. In 2017, Facebook opened several of its libraries, pre-trained models and data sets including its Caffe2 and PyTorch frameworks (Arakelyan 2017). Microsoft’s CNTK, Baidu’s Warp-CTC, and Amazon’s DSSTNE (the AI framework that powers its product recommendation system) are now all freely available and can be used to produce industry-grade AI (Stone 2016; Finley 2016). In addition, there are many other open-source AI projects not derived from the tech giants (Harvey 2017). Almost all AI projects today rely on such open-source toolkits. This is a significant milestone, we suggest, for AI becoming part of the general conditions of production.

means of cognition

53

Do such open-source projects mean AI will follow a path beyond the control of giant corporations? To answer this question, it is useful to remember the history of so-called free and open-source (FOSS) programming. Free software advocates such as Richard Stallman called for the non-commodification of software in the 1980s, but were largely defeated by the business-friendly open-source movement, championed not only by Linus Torvald, inventor of Linux, but also by the likes of Tim O’Reilly, CEO of O’Reilly Media (Liu 2018; Halliday 2018). As 2020 approaches, ‘open source’ is a buzzword for the business press and major IT corporations have shifted from seeing the open-source community as dangerously subversive to viewing it as a source of robust no-cost programming, a potential recruitment ground, and a strategic site for attracting users to their platforms (Weber 2004; Söderberg 2008; Tozzi 2017). Indeed, some open-source projects are dominated by contributions from employees of companies who use those projects. In the case of Linux, 2017 saw ‘well over 85 percent of all kernel development … done by developers who are being paid for their work’ (Corbet and Kroah-Hartman 2018: 15).

The example of Google’s Android operating system illuminates open-source’s shift. Android was released in 2008 by Google on an ‘open’ basis to challenge Apple’s domination of the smartphone market. While Apple’s iOS remains exclusive to its iPhone, Android, in the hands of Samsung, has since 2017 become the globally dominant smartphone operating system. But what has Google gained from this give-away? Initially, it was a defence to avoid Google’s search engine being cut out of an Apple-dominated mobile phone world. Subsequently, as Android itself rose to ascendancy, Google gained a widely-used platform. Google has now instituted ‘closed source creep’ in which ‘an open source base [is] paired with key proprietary apps and services’ (Amadeo 2018). Google has rendered more and more of the apps customers expect to find on an Android phone inaccessible not just to other operating services, but even to developers making versions of Android that Google does not control, such as the free Android clone Replicant. Although ostensibly open-source, in practice Android largely operates as an annex of Google’s larger data-harvesting operations, which in turn sustain its massive advertising revenues and training of its ML systems. In 2018, European Union antitrust regulators fined Google a record \$5.1 billion for abusing its power in mobile phone markets, declaring that ‘Google has used Android as a vehicle to cement the dominance of its search engine’ (Satariano and Nicas 2018).

54

inhuman power

Large corporate AI producers can thus not only coexist with, and indeed benefit from, open-source AI development, but can actually weaponize it against competitors (Vorhies 2016b). Google’s TensorFlow can be run on competing (non-Google) clouds such as Amazon Web Services or Microsoft Azure, but this is likely due to Google being the latecomer to the cloud game. It is now vigorously attempting to take a slice of the market. It is possible that a Google dominant in AI could cajole or coerce developers onto the Google cloud, and TensorFlow could become ‘the Android of AI’ (Gershgorn 2015). Further, as the tech companies themselves admit, open sourcing software can lead to accelerated development and improvements beyond what a single team at a company could accomplish. As one commentator observes:

Free software is good business for these companies, exactly because it allows more people to develop AI. Every big tech company is locked in a battle to gather as much AI talent as possible, and the more people flooding into the field the better. Plus, others make projects with the code that inspire new products, people outside the company find and fix bugs, and students are being taught on the software in undergrad and Ph.D. programs, creating a funnel for new talent that already know the company’s internal tools. (Gershgorn 2018)

Open-source AI projects thus act as ‘on-ramps’ to the proprietorial infrastructures of large AI companies (Asay 2017). Indeed, corporate encouragement of open-source AI is now so high as to compel critics of digital capitalism to worry not only that corporations will quash open-source AI, but also that their encouragement of it will produce malign results, such as the ‘deep fake’ pornography built with Google’s TensorFlow (Gershgorn 2018). So symbiotic have platform capitalism and open-source AI become since 2010 that both critiques may turn out to be true, with the commercially successful products of open-source developers gradually being consolidated under the aegis of large technology companies, while a shadow world of amateur dark-side AI is created on freely available corporate tools. These capital-open-source relationships typify what Paolo Virno terms ‘the communism of capitalism’ (2004: 110): corporations actively fostering a bottom-up, diversified and often free production of goods, and then harvesting this fecundity by commodifying its most successful fruits.

means of cognition

55

Such a strategy is consistent with the AI industry dynamics we have already reviewed. Major AI developers, themselves the direct and indirect beneficiaries of government-supported AI research, are both supplying AI capacities to other businesses and fostering the growth of large open-source communities. For capital as a whole, this means that AI-driven business analytics, managerial tools and production automation will become increasingly available, supplied from a combination of large cloud computing platforms and edge computing devices. For end-users, apps and products will increasingly integrate AI functions, and tools for making AI will be increasingly available and easy to use. Such a ‘democratization’ of AI will be entirely consistent with the reaping of massive profits by the major oligopolists of the AI industry – just as the production of earlier generations of general conditions of production, such as railways and telecommunications, created the fortunes for the corporate producers of such infrastructural technologies. If AI becomes generally available, it will still remain under the control of these capitalist providers.

THE SMART CITY, THE INTERNET OF THINGS,AMBIENT INTELLIGENCE

‘Democratization’ programmes are not the only way AI might be made generally available. Three other topics favoured by high-tech capital illuminate this possibility from different angles: the Internet of Things (IoT), the smart city and ambient intelligence. The IoT can be simply defined as the ‘pervasive deployment of [networked] smart objects (Kopetz 2011: 307). While the internet is commonly understood as a technology for communication between humans, the IoT is a hypothetical or emerging internet comprised of machine-to-machine communications, empowered by technologies such as radio-frequency identification (RFID).⁶ The term IoT was coined by Kevin Ashton (2009), who believes ‘[w]e need to empower computers with their own means of gathering information, so they can see, hear and smell the world for themselves, in all its random glory. RFID and sensor technology enable computers to observe, identify and understand the world – without the limitations of human-entered data.’ The goal of the IoT is for machines to dispense with human intermediaries and to communicate and act intelligently in the world on their own. Recapitulating Microsoft’s thesis about the fundamental constraint posed by human cognitive limitations, some analysts suggest that such a machinic population will require AI to be functionally realized: ‘the flood of data that comes from IoT devices … [has] limited value without AI technologies that are capable of finding valuable insights in the data’ (International Data Corporation 2016).

56

inhuman power

The smart city is a term lacking a consensus definition (Cocchia 2014), but visions of it provide another way of thinking about AI’s capitalist near future, focused on issues of urban development. Most conceptions of the smart city agree on at least one criterion: the presence of ‘pervasive and ubiquitous computing and digitally instrumented devices built into the very fabric of urban environments’ (Kitchin 2014: 1). These may include sensors and cameras hooked up to various actuators and processors which, through machine-to-machine communication, automatically optimize traffic, maintenance, energy distribution or various other urban flows in such ways that social life is improved. The smart city is thus a particularly urban manifestation of the IoT, but with a tacit political agenda of placing urban development increasingly in the hands of large AI-capitalists. As two critics put it, smart city discourse evinces a ‘free-floating utopianism about governance as a machine that would go of itself’ (Sadowski and Pasquale 2015). Google is, perhaps unsurprisingly, involved in smart city endeavours; as we discuss in Chapter 2, its project to establish a smart neighbourhood on Toronto’s waterfront was met with local resistance throughout 2018, although its fate remains uncertain.

The so-called ‘democratization of AI’, the smart city and the IoT are all different ways of expressing the notion of ambient intelligence, or ‘electronic environments that are sensitive and responsive to the presence of people’ (Aarts and Encarnação 2006: 1). The goal in the ambient intelligence paradigm is a situation in which ‘devices operate collectively using information and intelligence that is hidden in the network connecting the devices. Lighting, sound, vision, domestic appliances, and personal health care products all cooperate seamlessly with one another to improve the total user experience through the support of natural and intuitive user interfaces’ (Aarts and Encarnação 2006: 1). The goal is that rather than a user responding to their environment, user and environment are to be engaged in an ongoing bidirectional process of interaction (Aarts and Encarnação 2006: 11). There will be a ‘new intelligent intermediary layer between people and systems’ (Panetta 2017) and the environment will become ‘proactive’ in its interactions with users (Aarts and Encarnação 2006: 11).

means of cognition

57

Generalized ambient intelligence is one way AI could become part of the general conditions of production. If achieved, it would constitute a radical change to the technological milieu of capital, particularly if the AI which becomes ambient has ML capacities for perception and cognition. What would it be like if not only human knowledge and skills were transmuted into dead labour, but if dead labour gained the fundamental capacities for perceiving and cognizing that humans have historically monopolized? Perception and cognition would, like electrification, become ubiquitous and mundane properties of things in general. Without any claims to predict the future, we can look at some existing AI applications that currently function as fixed capital and imagine a situation in which their use becomes part of the general conditions of production.

One such application is the chatbot. Chatbots are ‘any software application that engages in a dialog with a human using natural language’ whether textual or auditory (Dale 2016: 813). As Helen Hester (2016) notes, these apps, intended to outsource to machines aspects of clerical, administrative and communications work, ‘represent, in many respects, the automation of what has been traditionally deemed to be women’s labour’ – a point manifest in the names and voices of Alexa, Cortana, Siri – technologies that ‘do gender’. Today, chatbots are predominantly textual, but Google’s Duplex (under development as of October 2018) has demonstrated uncannily human-sounding verbal conversation skills in limited settings. Some observers have even objected to its all-too-human ‘“um” and “ah” sounds’ and demanded that it announce itself as an AI (R. Metz 2018). According to one study, 36 per cent of businesses already employ chatbots and an additional 44 per cent expect to by 2020 (Oracle 2016). Chatbots can be simple rule-based programs, but to achieve more robust functionality and ease of use, modern systems usually employ ML to learn from experience and natural language processing to converse more easily. Google’s Duplex, for example, is an ML system built on a recurrent neural network (Leviathan and Matias 2018). Industry analysts expect that soon chatbots ‘will use AI to manage unstructured data and complex tasks’ (Panetta 2016). Chatbots are often embedded in messenger programs such as Facebook Messenger and WeChat and are most often employed in customer service applications. In its first year of opening up chatbot functionality, Facebook had more than 33,000 active bots on Messenger (Vr 2016).

58

inhuman power

The developers of Google’s Duplex hope that their creation will achieve the ‘long-standing goal of human-computer interaction’ of ‘making interaction with technology via natural conversation a reality’ (Leviathan and Matias 2018). This neatly sums up the purpose of the chatbot: to act as a new interface that replaces previous customer-facing business elements, such as web stores and human technical support, with automated natural language conversation. The chatbot is intended to make online transactions as intuitive and simple as in-person transactions. Consultancy firm Gartner anticipates that ‘[i]nteracting with chatbots won’t require any particular set-up; the technology will simply understand and do as the human asks’ (Panetta 2016).

Chatbots have not only found use in customer service, however. As Lebeuf, Storey and Zagalsky have shown, software developers have been creating their ‘own breed’ of chatbots to ‘reduce collaboration friction’ in the workplace (2017: 2). As the authors demonstrate, developers use chatbots to promote the functionality of their group by coordinating schedules and tasks, promoting adherence to group norms and roles, demarcating roles, responsibilities and expertise, as well as monitoring and promoting cooperation and trust (2017: 3–4). They also use chatbots to curate large information flows and share knowledge and skills (2017: 5). In so doing, they are offloading chunks of their social activities to these bots which convert various tasks and processes into dialogic form.

2018 was a ‘milestone year’ for chatbots in terms of both technical advances and business applications (Seth 2018). Yet, due to the domain-specific content required for each chatbot, these systems suffer from the same problem of heavy customization workload that faced GOFAI expert systems. However, research is underway on the possibility of ‘bootstrapping’ or automating the production of chatbots in new fields of application using ML trained on data available in that domain (Babkin et al. 2017). If such work is successful, we can expect chatbots to proliferate further and approach the ubiquity we have posited for AI as a general condition of production.

We can also consider another existing AI application which is expected to become widespread. In 2016, Amazon opened its first branch of Amazon Go, an automated convenience store, in Seattle. As of early

means of cognition

59

2019, a tenth branch is under construction. Go, billed by Amazon as ‘just walk out shopping’, uses ‘computer vision, deep learning algorithms, and sensor fusion’ to dispense with cashiers and check-out lines (Amazon 2016). Essentially, Amazon Go relies on an array of devices for machine perception, the data from which is processed and synthesized by AI. Customers sign in to the store with the Amazon app and their actions within are tracked such that whichever items they pick up are automatically noted by the store, and upon walking out, their account is charged. Despite Amazon’s automation rhetoric, as of 2019, Go stores still involve human fixers and overseers, much as autonomous vehicles do (Del Ray 2017). Melville (2017) suggests that rather than seriously delving into the low-margin grocery industry, Amazon is merely using it as a venue for testing out a new method of payment, i.e. a new interface for retail transactions. Through pervasive AI-powered machine perception, Amazon Go transforms the retail transaction from a human-to-human interaction, in the case of a cashier, or a human-to-machine interaction, in the case of self-checkout, into a commodity-to-machine interaction which remains invisible to the human customer.

Extrapolating from these examples, we can begin to imagine what it would be like if capacities for machine perception and cognition were generalized throughout society and thus became a general condition for production. Like electricity, AI perception and cognition could be put to many uses. In the cases discussed above, they are employed with the obvious goal of the replacement of human labour – we turn to the topic of AI effects on employment in the next chapter. But if we approach these cases from the perspective of limited human cognitive capacities as fetter posed by Microsoft, we can see how they might be more generally applied. In the cases of AI-enabled chatbots and automated retail, AI cognition and perception are being deployed to streamline processes of social interaction by converting them into simplified, easily digestible forms administered by machines. Complex workplace social interactions among software developers and customer service transactions are thus reduced to dialogue with a chatbot which facilitates the process. In the case of Amazon Go, the retail transaction disappears altogether from human phenomenology, transformed by the integration of various data streams.

In both cases AI acts as an interface which simplifies a complex situation. This is nothing new for the capitalist mode of production. As Vincent Manzerolle and one of us (Manzerolle and Kjøsen 2015) show in the context of near-field communication (NFC) devices, capital has enthusiastically embraced technologies which simplify and speed up transactions. Excited by AI’s capacities for further simplification and speed, the consulting firm Accenture (2017) has declared that AI will be the next user interface (UI), replacing the graphic user interface that is ubiquitous on our screens today. One of us has pointed out that analysts at Accenture deploy the unusual term ‘curation’ to describe the novelty of AI when it is considered as an interface (Steinhoff 2019a). Accenture declares that ‘at the height of sophistication, AI orchestrates. It collaborates across experiences and channels, often behind the scenes, to accomplish tasks. AI not only curates and acts based on its experiences, but also learns from interactions to help suggest and complete new tasks’ (Accenture 2017: 20). The vagueness of this description confirms the imagined widespread applicability of the function of curation – an intelligent, adaptive technique of cognitive automation. AI functions as an interface which not only represents some information to a user, but by perceiving and cognizing, actively gathers, processes, reveals and hides information before and while the user acts on it. This is the imagined function of the ‘new intelligent layer’ posited by ambient intelligence advocates (Panetta 2017) and Kelly’s ‘cheap, reliable, industrial-grade digital smartness’ (2014).

60

inhuman power

Recall that, for Marx, infrastructure functions to ‘facilitate circulation or even make it possible at all, or … increase the force of production’ (Marx 1993: 530–1), while the function of the general conditions as such is to give the mode of production an elasticity, i.e. a capacity for expanding the volume and velocity of production. Roads and digital networks facilitate circulation. Education systems increase the force of production by producing and distributing knowledge and ideology. In the context of capital’s collision with the fetter of limited human information-processing capacities, infrastructural AI is being positioned to curate information flows, saving the human cognitive apparatus for whatever machines cannot yet handle. The retail transaction, essential to capital, yet annoyingly (for capital) consumptive of time, thought and wages, can be etherealized and accelerated with Amazon Go’s AI. This transaction is effectively being curated out of existence by Amazon, and while human agents of Amazon directed this process, more advanced AI systems may develop and enact their own processes of curation, or these might arise from the interaction of multiple AI systems. It is not possible to enumerate here all the ways an infrastructure of AI could optimize circulation and/or increase the force of production. Already in the works are systems such as SIX’s (n.d.) ‘complete infrastructure for ecommerce over the Internet of Things’, which aims to link smart home appliances (such as refrigerators which perceive when supplies are running low) with market information, retailers and delivery services to completely eliminate shopping from the purview of human cognition and action. There are certainly many other ways in which capital’s valorization and realization could be powered-up by offloading cognitive and perceptual capacities to the environment at large.

means of cognition

61