From the Series: The Infrastructure Toolbox
In 2014, the Evening Standard newspaper listed Demis Hassabis as London’s third most influential person. After completing his PhD in neuroscience, Hassabis, a former computer game designer, world gaming champion, and chess master, started a company called DeepMind. Three years later, Google paid £400 million for the company. DeepMind had no products, but the company built neural networks that learned to play early-1980s Atari computer games without any training (see Mnih et al. 2015). This game-playing performance attracted the interest of Google, where Hassabis’s work is reportedly already shaping search engine results.
Game-playing neural networks might seem a long way from roads, water, energy, transport, and other infrastructures of modernity. Or does it? Two considerations suggest proximity between infrastructural modernity and projects like DeepMind:
1. Joining Hassabis at the top of the list of influential Londoners was Mayor Boris Johnson, the elected official responsible for the city’s infrastructure, and George Osborne, the government minister responsible for national finance and economic management. What might connect these three men? Might we say that after the infrastructures of global cities and state modernity comes the promise of mindful infrastructures? By mindful, I mean the attempt to lend coherence to something that otherwise would threaten to appear distracted, splintered, fragmented, or somehow aleatory. For instance, it might be worth thinking about how Hassabis’s work on personality and the hippocampus might inform the operations of search engines (Hassabis et al. 2013). A problematization, “a kind of general historical and social situation—saturated with power relations” (Rabinow 2003, 19), might be taking shape there.
2. DeepMind is not an isolated case of the melding of pattern recognition with information infrastructures. We could turn to IBM’s Watson, an enterprise-grade artificial intelligence project that became champion of the television show Jeopardy in 2011. Watson has since been networked and infrastructuralized as a cognitive computing assemblage that has, with the help of $1 billion in investment and a footloose global team of technical and marketing personnel, rapidly penetrated elite medical institutions and wedded itself to insurance, higher education, health industry, scientific fields (notably genomics), and cooking. The financialized leveraging of these approaches is also well evidenced in the case of Hewlett-Packard’s Autonomy. Autonomy, one of the United Kingdom’s largest and most profitable software businesses, was bought by Hewlett-Packard (HP) for £7.4 billion in 2011 for its pattern-recognition techniques, which were branded as “meaning-based computing.” While HP soon wrote down the value of its acquisition by more than £5 billion, alleging accounting improprieties, the company integrated key components of meaning-based computing into its utility, healthcare, transport, call-center management, data management, operations management, patient records, legal support, security, and intelligence products.
DeepMind, Watson, and Autonomy are all highly mindful in their promise to reintegrate the dispersed, forgotten, or contradictory experiences of infrastructure. They abound in references to cognition, meaning, perception, sense data, hearing, speaking, seeing, remembering, deciding—and surprisingly—imagining and fantasizing. They polymorphously figure infrastructural reorganization around the ideal of something like pattern recognition or cognitive awareness. Their modeling practices are no longer the statistical rendering of numbers in the hands of government, science, or commerce (Porter 2008) or the interweaving of classification systems and things that grew into a forest of operational standards over the course of the last century (Bowker and Star 1999). DeepMind, Watson, and Autonomy each address the relation between humans and infrastructures, not so much as a matter of imagining, practice, configuration, or repair, but as a competitive cognitive challenge. They present problems of seeing, hearing, checking, and comparing as no longer the province of human operators, experts, professionals, or workers seeking to navigate and finesse the constraints, limitations, breakdowns, and vicissitudes of infrastructures, but as challenges set for an almost-cyclopean cognition to reorganize and optimize in ongoing competitive experimentation. The initial training of DeepMind to play Atari games or Watson to win Jeopardy indexes this orientation toward challenges amidst competition, disconnectedness, and disparity.
Would it be fair to say that these kinds of cognitive infrastructures, with their appetite for data and their ambition to reorganize other infrastructures, emanate from the techno-cosmological imaginaries of Silicon Valley engineers and the like? Yes and no. On the one hand, the centers of calculation that Google or HP make and manage are effectively global assemblages (Ong and Collier 2005), with specific administrative, commercial, engineering, and scientific apparatuses and regimes of value. Undoubtedly, they powerfully subduct existing infrastructures. On the other hand, the increasing mindfulness of the infrastructures under construction at companies like IBM and Google predicate a certain reconcatenation of the world, no longer found in the mobile train of experience that moves through streets, houses, factories, and offices, but instead in the relations mindfully discerned in streams of data. Contemporary infrastructures might change shape as infrastructural mindfulness gathers scattered, partial, and disaggregate worlds, themselves products of infrastructural splintering and the device-specific intensities of “knowing capitalism” (Thrift 2005).
Bowker, Geoffrey C., and Susan Leigh Star. 1999. Sorting Things Out: Classification and Its Consequences. Cambridge, Mass.: MIT Press.
Hassabis, Demis, R., Nathan Spreng, Andrei A. Rusu, Clifford A. Robbins, Raymond A. Mar, and Daniel L. Schacter. 2013. “Imagine All the People: How the Brain Creates and Uses Personality Models to Predict Behavior.” Cerebral Cortex.
Mnih, Volodymyr, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, et al. 2015. “Human-Level Control through Deep Reinforcement Learning.” Nature 518, no. 7540: 529–33.
Ong, Aihwa, and Stephen J. Collier, eds. 2005. Global Assemblages: Technology, Politics, and Ethics as Anthropological Problems. Malden, Mass.: Blackwell.
Porter, Theodore M. 2008. “Locating the Domain of Calculation.” Journal of Cultural Economy 1, no. 1: 39–50.
Rabinow, Paul. 2003. Anthropos Today: Reflections on Modern Equipment. Princeton, N.J.: Princeton University Press.
Thrift, Nigel. 2005. Knowing Capitalism. London: Sage.