人工智能和资本主义的未来

THE AI INDUSTRY AND THE OLIGOPOLISTS OF MACHINE INTELLIGENCE

While corporate interest in the actual and potential uses of the new AI are manifold, ranging from retail sales to entertainment and industrial production, the actual production of AI systems is a central concern for a more limited circle of high-tech companies. We refer to this complex as ‘the AI industry’, distinct from the broader field of commercial AI applications. While business-oriented publications continually remind us that AI will ‘revolutionize’ capitalist production (Columbus 2016), our analysis suggests that such a transformation, if it occurs, is still in its earliest phases. Instead, we see AI as one emerging industry whose influence is tied up with that of other emerging technologies and is as yet difficult to ascertain with certainty. Although business interest in AI is high, outside the AI industry this does not entail high levels of actual investment in the technology. A 2017 survey of attendees at an applied artificial intelligence conference concluded that ‘AI adoption … remains low with the majority of major success stories coming only from the largest tech players in the industry’ (Rayo 2018).

AI development first appeared as a distinct industrial sector in the 1980s. This first era of the AI industry was based around GOFAI expert systems. During this era, the AI industry consisted of a few small companies which produced systems as means of production for, and typically in cooperation with, their corporate customers. In some cases, large firms established internal AI departments to develop proprietary expert systems. Such systems required a considerable degree of specialization, had extremely narrow fields of application and required a lot of labour to produce and update. While attempts were made to develop ‘generic’ expert systems which could be applied to any field, they ultimately failed (Roland and Shiman 2002: 205). The commercial craze for these systems subsided in the 1990s, but around the same time the ML approach gained traction in academia and, during the 2010s, returned AI to the commercial realm, propelled by advances in computing power and improved learning algorithms. By 2017, The Economist (2017a) was proposing a shortlist of domains in which ML’s power to ‘sift through data to recognize patterns and make predictions without being explicitly programmed to do so’ was becoming commercially important. It is worth surveying a few.

32

inhuman power

The ML-based AI industry is much more diverse than the first era of expert systems; this is one reason why advanced capitalism has recently contracted a serious bout of AI fever. The Economist (2017a) has been enthusiastic about the prospects of targeting online advertisements and product recommendations; the creation of virtual personal assistants and of augmented reality systems; and autonomous vehicles. As of early 2019, some of these were already highly advanced, while others only incipient. Algorithmic targeting of advertisements and recommendations has been a foundation of digital Web 2.0 enterprises for over a decade. Digital personal assistants, such as Apple’s Siri, Amazon’s Alexa or Microsoft’s Cortana, are gradually becoming commonplace. Augmented reality (AR) products, overlaying physical reality with a mesh of virtual images and information, are only beginning to be sold as commodities or distributed as free vehicles for in-app purchases and data mining. Games such as Pokémon Go and other mobile apps are testing the AR waters, while further frontiers, such as medical applications, are being actively researched by companies such as Google, Apple and Microsoft. Perhaps the biggest prize for the commercial use of ML, but also its most daunting challenge, is the creation of self-driving cars and trucks, a ‘moonshot’ that has attracted leading information companies such as Google and Baidu, established auto-industry giants such as Ford, General Motors and Daimler, and upstart entrants such as Uber and Tesla, all racing to transform capitalism’s entire transportation sector.¹

AI industry enterprises build ML technologies, often initially for use in their own business operations, but also as commodities for sale or rent, or as a ‘free’ service. They produce commodities for both of the major ‘departments’ into which Marx divided society’s total product and its total production process: (Department 1) means of production, i.e. commodities intended for productive consumption; (Department 2) means of subsistence, i.e. commodities destined for individual consumption (1992: 471). Some commentators on ML have suggested that, neatly corresponding with these two departments, there will be ‘two AIs’: one for business applications, the other for consumer devices (Economist 2017a). In Department 1 we find examples like SAP’s

means of cognition

33

HANA, a ML-powered cloud database platform that enables behemoths like Walmart to monitor their entire organization’s functioning in fine-grained, real-time detail (Ruth 2017), and Andrew Ng’s start-up Landing.ai (founded in 2017) which aims to totally overhaul industrial manufacturing by providing ‘AI-powered adaptive manufacturing, automated quality control, predictive maintenance, and more’ (Landing n.d.). In Department 2, examples include various consumer commodities like Amazon Home and similar devices. Marketed as a ‘smart speaker’, Home is a user voice interface to the Alexa digital personal assistant that enables a variety of home automation and organizational tasks. AI is also found in other smart devices like phones and TVs and is also ‘given away’ as a component of free product-services such as Facebook, Twitter or YouTube where ML-based recommender systems curate timelines and give users suggestions on what to watch or listen to next. In turn, these systems gather customer data to fuel advertising revenues. However, as we will see, production of both Department 1 and Department 2 AI is often dominated by the same oligopolistic corporations, and may also be interconnected in a variety of ways, including the use of shared cloud computing facilities.

From 2015 on there has been a rapid escalation of corporate investment in AI research, venture funding of ML start-ups, and competitive hiring of AI talent as well as lots of acquisition and merger activity. Measuring the scale of this activity is difficult. According to one analysis, the AI industry had a revenue of \$126 billion in 2015 and is projected to grow to \$3,061 billion by 2024 (Statista 2016: 9), but another reckons worldwide spending on AI stood at only \$19.1 billion in 2018, an increase of 54.2 per cent over 2017, and predicts it will reach a mere \$52.2 billion by 2021 (International Data Corporation 2018). The Economist (2017a) calculates that in 2017 companies globally spent around \$21.3 billion in mergers and acquisitions related to AI – 26 times more than in 2015. While such conflicting estimates (often manifestly driven by the self-interest of AI vendors and business consultancies) are confusing, it is clear that AI has seized the imagination of advanced capital’s representatives (see also Press 2018). As The Economist (2017a) puts it, ‘Fueled by rivalry, high hopes and hype, the AI boom can feel like the first California gold rush.’

Corporate competition for ML experts is ferocious. One study, based on LinkedIn profile data, puts the number of PhD educated people ‘capable of working in AI research and applications’ at 22,000, with only 3,074 currently looking for work (Gagné 2018). Demand far exceeds supply (Economist 2017b). US information capitalists are in competition both with new contenders – such as major auto companies with autonomous vehicle projects – but also now with China’s tech companies, some of which have set up subsidiaries in Silicon Valley. As hiring top talent is seen as crucial for the success of AI-capital, this competition has ‘set off a trend of firms plundering academic departments to hire professors and graduate students before they finish their degrees’ and created an atmosphere in which job fairs resemble frantic ‘Thanksgiving Black Friday sales at Walmart’ (Economist 2017b). This competition for ML talent also means that wages are high.

34

inhuman power

A recent New York Times article reports that ‘Typical A.I. specialists, including both Ph.Ds fresh out of school and people with less education and just a few years of experience, can be paid from \$300,000 to \$500,000 a year or more in salary and company stock, according to nine people who work for major tech companies or have entertained job offers from them’ (Metz 2017b). When Google acquired DeepMind in 2014, it paid \$650 million for a company of 50 employees; in 2016, the lab’s ‘staff costs’ alone, as it expanded to 400 employees, totalled \$138 million, an average of \$345,000 an employee. In the light of such figures, it has been suggested that ML experts are ‘the new investment bankers’ (Shead 2017). The rewards are even higher, of course, for executives with experience managing AI projects. In a court case against Uber over ownership of autonomous car technologies, Google revealed that one of the leaders of its self-driving car division took home over \$120 million in incentives before jumping ship to join their competitor. However, even fresh graduates with skills in ML may make ‘in excess of £100,000 and sometimes up to £1 million’ while still in their mid-twenties (Shead 2017).

The AI industry is international in scope. Between 2016 and 2018 it became widely recognized as a critical axis of technological competition between the United States and China, particularly given its potential for military application in an era of growing tensions. Important Chinese AI developers include its largest search engine corporation, Baidu, and ecommerce giant Alibaba (K-F. Lee 2018). Other important national sites for the AI industry include Canada, Israel and the United Kingdom. However, nearly all assessments suggest that the United States is the leading location (Jang 2017; Rapp and O’Keefe 2018; Fabian 2018). By one estimate, which surveyed over 3,000 companies around the world involved in aspects of AI development, 40 per cent are in the

means of cognition

35

US (Fabian 2018). Six, however, are preeminent: Alphabet (Google’s parent company), Amazon, Apple, Facebook, IBM and Microsoft. These companies all exemplify what Tarleton Gillespie (2010) and Nick Srnicek (2016) respectively describe as ‘platforms’ or ‘platform capitalism’, a key feature of which is the digital gathering of big data generated by customers, be they users of search-engines, social media networks and video or music streaming services, or computer software or retail consumers. Access to such troves of data makes platform firms favourable sites for training ML systems.

IBM

Amongst these, IBM is in many ways an outlier, even though ‘Big Blue has a long record of interest in AI, stretching from its researchers’ involvement in the famous 1956 Dartmouth workshop to the triumph of its chess-playing Deep Blue over world champion Garry Kasparov in 1997 and its AI Watson, whose 2011 victory over human competitors in the television quiz show Jeopardy made it briefly the public face of the new generation of AI. Yet despite IBM’s \$15 million investment in the system, Watson has subsequently had only limited commercial success. While it has been described as ‘one of the most complete cognitive platforms available’ (Kisner, Wishnow and Ivannikov 2017: 1), and has been applied eclectically to commercial ventures in fields from fashion to telecommunications, IBM’s major emphasis was on potential uses in the highly profitable medical and health insurance sectors. In 2018, however, the company laid off many of the staff in this key division and announced it would be seeking new areas of focus. It is uncertain how far this setback was the result of technological failures and how much was due to the rigidities of IBM’s organizational practices (Strickland 2018). IBM is likely hampered in its AI efforts due to not possessing the large proprietary pools of big data necessary for training ML systems; instead IBM has to acquire it, expensively, by buying up smaller firms engaged in medical research and data collection (Kisner, Wishnow and Ivannikov 2017: 19–20). The other major US AI producers, however, do not suffer from this problem.

Alphabet (Google)

Alphabet has been harvesting user data and applying it to advance their AI projects for years, first by algorithmically improving search patterns and matching them with ad placements, and then using similar methods for categorizing, filtering and recommending video content on YouTube or predicting which apps users of its Android mobile phone operating system would purchase. Alphabet’s Google Brain unit is widely seen as the leading corporate ML research group. Between 2014 and 2018 Google bought up no less than 12 AI-related companies (Patrizio 2018), the most notable being DeepMind, which made the ML system AlphaGo that in 2016 scored an uncanny victory over the reigning human Go world champion, thus supplanting Watson as the poster-boy for AI. Such research connects not only with Google’s algorithmic online services and its Google Home devices but also with its Waymo autonomous vehicles unit and suite of robotics-related companies it acquired in the early 2000s. The development of AI is an endeavour fervently advocated by Google’s owners Sergey Brin and Larry Page as well as the transhumanist thinker Ray Kurzweil who is their ‘director of engineering’ (Simonite 2017); the combination of vast funds, deep expertise and ideological commitment places Google in an exceptional position in commercial AI research. Other US platform capitalists are, however, following similar paths.

36

inhuman power

Facebook and Amazon

Algorithmic analysis and prediction have been central to the success of Facebook in plotting the ‘social graph’ of users’ interests and interrelations which drives its massive online advertising revenues. Facebook AI Research (FAIR) has four AI laboratories around the world, is active in conducting cutting-edge AI research, and has made several AI-related acquisitions, such as the company Ozlo, which builds virtual assistants (Patrizio 2018). Amazon’s development of ‘recommendations’ for customers across its escalating retail and logistics operations relies on the algorithmic analysis of vast volumes of consumer data, now increasingly integrated with the operations of its huge and partially robotized warehousing and order fulfilment systems. Amazon has a number of specific AI-based products, including the Echo ‘smart home’ system; its digital personal assistant, Alexa; Lex, a business version of Alexa; Polly, which turns text into speech, and Rekognition, an image recognition service. In addition, ML permeates the suite of services in the Fulfilment by Amazon programme the retailer offers to third-party sellers.

means of cognition

37

Microsoft and Apple

Although of an older generation of IT companies than Google, Facebook and Amazon, Microsoft draws not only on decades of software development experience, but also on the records of interaction with millions of computer users for projects such as Cortana (the digital assistant bundled with its software), and more and less successful chatbot ventures, ranging from the catastrophic Tay (which machine-learned from online conversations to be a racist, sexist Nazi that had to be terminated with extreme prejudice) to the more innocuous teenager-imitating Zo. Finally, Microsoft’s long-time competitor, Apple, which initially seemed the least AI-avid of the major platform capitalists, has also entered the arena in 2016–17, with the ongoing development of its digital personal assistant Siri and the creation of FaceID, a facial recognition security system for iPhone users. In 2018, it dramatically poached Google’s AI chief (Patrizio 2018).2

Beyond the Tech Giants

Around the top-rank tech corporations are clustered many smaller start-up companies that are attempting to carve out niches in specialized branches of the AI industry, ranging from biotechnology, to farming, education, merchandising and surveillance (Zilis and Cham 2016; Patrizio 2018). This scene is ever changing. It is likely that, in a pattern familiar from previous cycles of IT development, many of these companies will flare up and burn out, with the successes likely to be acquired by AI industry giants: ‘115 of 120 AI companies that exited the market in 2017 did so by acquisition’ (Patrizio 2018). Companies producing specialized hardware for AI systems also continue to appear in the shadows of dominant hardware firms like Nvidia and Intel. Venture capitalists invested \$1.5 billion in hardware start-ups in 2017, twice as much as two years previously; new companies such as Cerebras, Graphcore and Cambricon have each attracted over \$100 million in speculative funding. Such start-ups ‘are racing toward one of two goals: Find a profitable niche or get acquired. Fast’ (C. Metz 2018).

STATE ACTORS: AI SUPERPOWERS

Another crucial actor in AI development is the state. Marx never completed a systematic study of the capitalist state, and attempts by his followers to do so have stirred some of the most complex debates in Marxist theory.³ However, one function Marx clearly did ascribe to the state in capitalism was the creation of certain general conditions of production (Marx 1990: 506; 1993: 530) such as infrastructures, which could then be transferred to private capital once they became profitable ventures. Such privatized industries in turn provide capitalist states with technological powers exercised in the name of national security against rivals: this is the dynamic of military-industrial complexes. This path of state-capital interactions is classically demonstrated by the development of digital technologies in the US. Computers and networks were incubated in the military-corporate-academic wing of the Pentagon before passing into general commercial use (Edwards 1996; Mazzucato 2013). The most celebrated, but by no means singular, case is the funding of the internet’s creation by the US Advanced Research Project Agency (ARPA). Contrary to the libertarian self-presentation of digital capital, the creation of US high-tech industries depended on state-sponsored research, subsidization and contracts for the creation of the technologies which played a major role in America’s Cold War victory over its state-socialist opponents.

38

inhuman power

Such state-capital interactions in the development of digital technologies were widely adopted beyond the US in the era of globalization following the end of the Cold War. While many theorists, both bourgeois and Marxist, declared the decline of the nation state, in actuality, capitalist globalization unfolded through the mediation of nation states whose activities characteristically involved the support and subsidization of the digital industries and infrastructures on which competitive participation in the world-market depended (Schiller 1999; Powers and Jablonski 2015). The so-called ‘Washington consensus’ governing globalization masked mounting antagonisms, in particular between a Cold War-victorious United States and the defeated ‘post-socialist’ Russia and China, which were compelled to capitulate to or compromise with US-led capitalism. In 2008, when the Wall Street crash manifestly weakened the US as imperial hegemon, these conflicts emerged more sharply in a flare up of economic and military rivalries to which contending nation’s plans for AI are now integral.

US AI development was already built on the basis of state-initiated infrastructures and technologies; leaders of AI research, such as IBM and later Google (Nesbitt 2017), have been supported by US defence-related funding: according to one report no less than 16 US government agencies fund AI development (Fabian 2018).⁴ However, it was only in 2016 that the US Federal Government began formulating an overall National Artificial Intelligence R&D Strategic Plan. The Trump administration has named AI a national priority because of ‘its role in helping the U.S. lead in technological innovation as well as its role in information statecraft, weaponization, and surveillance’, while the President has made it clear that he will not ‘stand in the way’ of AI-capital by burdening it with regulations about the social or employment effects of AI (Future of Life Institute 2018). Amongst a variety of measures, the US Department of Defense’s announcement that it would invest up to \$2 billion over five years towards the advancement of AI was prominent: use of ML in surveillance and for the control of ‘swarming’ drones and other semi-autonomous weapons appear to be a high priority (Scharre 2018). High-tech workers at Google and some other Silicon Valley companies have revolted against participation in military contracts – an important instance of resistance to AI we discuss in Chapter 2. Despite these protests, however, the Pentagon made clear that this was only the first phase of an ‘AI surge’ (Seligman 2018) that would include other projects such as the \$10 billion to be spent over the next ten years for a special military cloud computing Joint Enterprise Defense Infrastructure (JEDI).

means of cognition

39

This sudden urgency about AI on the part of the US state is attributable to the emergence of a serious digital rival: China. In 2017, the People’s Republic announced its Next Generation Artificial Intelligence Development Plan, with initiatives and goals for R&D, industrialization, talent development, education and skills acquisition, standards, regulation, ethics and security. It envisages China becoming the ‘primary’ centre for AI innovation by 2030 (Dutton 2018). Given the immense gulf in technological development that separated China and the US even in 2000, this ambition seems staggering. What gives it some plausibility is both the pace of China’s economic growth since 2000 and changes in the nature of AI development. Kai-Fu Lee (2018), former head of Google China and now a champion of China’s AI programme, argues that the most recent generation of AI – ML in particular – has passed from the moment of watershed breakthroughs to one of innovative application. Lee holds that in this scenario, China has an AI advantage in so far as it possesses large numbers of software engineers (not necessarily of superstar quality, but highly proficient), a fiercely competitive digital capital sector, and huge quantities of data gathered virtually without restraint. He suggests that the United States and China will, by 2030, constitute a state level ‘duopoly’ in terms of control of the AI industry. While his account ends with benign hopes for cooperation, the underlying vectors of intense economic and military competition are all too visible in his account of a ‘new world order’ dominated by ‘AI Superpowers’ (K-F. Lee 2018).

40

inhuman power

Other states are understandably reluctant to acquiesce to this vision of an America-China AI duopoly. In 2018, the European Union acknowledged that fierce international competition demanded coordinated action for it to remain at the forefront of AI development and announced a joint ‘public-private’ programme aimed to increase investments in AI research and development by at least €20 billion by 2020 (Middleton 2018). A striking demonstration of the centrality of state policy for AI development, and a major reason for European corporations’ anxieties on this score, is the potential impact of the EU’s recent General Data Protection Regulation (GDPR). This legislation places limits on corporate data gathering that are stringent by comparison with the policies of the US and China. The GDPR is criticized by business lobbyists for cramping potential ML projects – a disturbing harbinger of the potential for AI-capital to normalize and require regimes of high-surveillance governance. In an attempt to staunch a ‘brain drain’ of AI talent to super-salaried positions with US and Chinese corporations, the EU also launched a multinational European AI institute – the European Lab for Learning and Intelligent Systems (ELLIS) – with centres in a number of European countries (Rankin 2018). Beneath this common European front for AI, however, national rivalries continue to roil, with the UK, France and Germany jockeying for position as the regional AI leader (Shead 2018).

The US, China and the EU are merely the largest contenders in a worldwide rush by states to attract AI-capital. Japan and Russia are also substantially subsidizing AI development; by mid-2018 more than 20 nations, ranging from New Zealand to Poland, Kenya and Tunisia, had formulated ‘artificial intelligence strategies’ (Dutton 2018). Deep anxieties underlie such mobilization. As we will see, at the level of individual corporations, AI-capital may tend to favour winner-take-all concentration of ownership and the creation of monopolies. A similar dynamic may emerge at the level of international state relations. One of the more alarming features of Lee’s predictions of a ‘bipolar’ US-China dominance of AI-capital is that ‘while AI-rich countries rake in astounding profits countries that haven’t crossed a certain technological threshold will find themselves slipping backward ... [into] a state of near total dependence and subservience’, frantically ‘trading market and data access’ for the use of AI facilities beyond their control (K-F. Lee 2018: Kindle Loc. 2759). This prospect, along with the yet blunter threat of AI military power, ensures that no state wants to be excluded from developing the emerging general conditions of capitalist production – and destruction.

means of cognition

41