Five’s heroes should care about its individual reward function versus the average of the team’s reward functions. We anneal its value from 0 to 1 over training’ (OpenAI 2018). Here cooperation results from the optimization of numerical values.
What about the capacity for abstraction or the ability to generate and grasp an ‘infinite variety of concepts and logical schemes’ (Virno 2004: 106)? The pioneers of deep learning address this directly when they describe the distinctive capacity of deep learning systems as ‘learn[ing] representations of data with multiple levels of abstraction’ (LeCun, Bengio and Hinton 2015: 436). Or as a tutorial from IBM puts it: ‘By building multiple layers of abstraction, deep learning technology can solve complex semantic problems’ (Sedlak 2016). This is not mere hype. Machinic language translation is now done with neural networks because the networks are able to find patterns and develop relevant linguistic concepts (rules) which produce better translations than any other automated technique. In addition, deep learning systems are immensely scalable, meaning that the more data that becomes available, the more the performance of these systems increases (Ng 2015). The performance of older algorithms plateaued at a certain amount of data, but, so far, deep learning has not. The epistemic, logical and semantic complexity that such systems will be able to grasp in the future is impossible to predict.
Finally, we come to affect, a capacity historically inimical to machines. While there is, as of 2019, no evidence of machines expressing affect, the ability of AI to capture and process data on the emotional states of human beings continues to improve. AI techniques are being deployed in the field of sentiment analysis, which seeks to computationally recognize subjective states including emotions and opinions in text (Pang and Lee 2004), in emotion recognition in speech (Morrison, Wang and De Silva 2007), and in recognition of facial emotions (Zhang et al. 2017). While Japanese companies are investing heavily in care robots, a wide variety of businesses are becoming interested in using AI to monitor and manipulate the emotions of their customers (Faggella 2018a). AI may not experience joy or sadness any time soon, but AI will likely be increasingly used by capitalist firms to add affective dimensions to diverse facets of production, circulation and beyond.
As these examples show, the capacities Virno reserves for a human general intellect are in the process of AI-driven implementation. What place then might human labour find within AI-capital, if not the one reserved by post-operaismo? A passing remark made by Marx is suggestive. In the ‘Fragment on Machines’, Marx imagined a highly automated workplace where ‘the human being comes to relate more as watchman and regulator to the production process’ (1993: 705). Here a human attendant watches over the lights-out factory, present only to press the stop button if something goes awry. The rarely quoted following sentence (which is interestingly bracketed) reads: ‘What holds for machinery holds likewise for the combination of human activities and the development of human intercourse’ (1993: 705). The term ‘intercourse’ was translated by Martin Nicolaus from the German ‘verkehrs’, but another translation would be ‘traffic’. As Matthew W. Bost shows, Marx used verkehr to refer to all kinds of intercourse including biological reproduction, linguistic and semantic exchange, commercial exchange, as well as ‘the exchange of weapons fire in war’ (2016: 338). Thus, here Marx seems to suggest that humans may also come to relate as mere watchmen to the development of their societies and the infrastructures (broadly understood) which underpin them. The offloading of cognitive capacities entailed in the establishment of the means of cognition could provide the means for capital to achieve such deep control.
66
inhuman power