Lesser Gods: Labor in the AI Labyrinth, Part 1
Advance Drafts #3
What follows is the first of three parts of an extensive draft on the AI supply chain, based on my own firsthand experience as a worker assembling models within the “AI factory.” This was originally intended as a simple, one-off article detailing the boom in employment in these new AI assembly jobs and tracing out the global division of labor through which models are produced. However, given the pace of development within the industry and the sheer amount of absurd misunderstandings about the nature of the technology as well as continual controversies over data centers, energy use, military applications, etc., I soon found myself having to write out various clarifications and add expansions to the original plan. The article quickly got out of hand and now sits at roughly 30,000 words, still growing. The first part, posted below, is roughly 13,000 words.
Rather than cutting it back down to a manageable size, I’ve instead tentatively decided to instead write it up further into a short book. I’ll then create a more polished article version, cut down from the larger content, to put forward the core argument and preview the content of the forthcoming book project. I’ll be posting sections of the initial draft here, before it is formatted into its proper book form with a linked narrative. While Hellworld carried on the project begun in Hinterland, which I hope to conclude with a third volume on the questions of partisanship and political organizing broached in my recent pieces for Ill Will and (forthcoming) Heatwave, this AI book is more of an aside. Think of it like a slim political pamphlet calling attention to a single industry, offering a brief bit of theory on its character, and posing some of the key questions for organizing in and around this emergent sector.
As always, please note that this is a rough draft. It includes typos, some awkward wording, and many extraneous bits that will eventually be refined or cut out. If citing it, please cite the forthcoming piece and don’t quote directly until that final version is released. Also, I’d strongly recommend taking a look at the special episode the excellent podcast Until Everyone is Free did on the industry in Kenya:
Introduction
Look at the computer in your hand, on your desk, hidden and humming in the depths of the server room. The sleek aluminum carapace. The glass screen like the jelly of an eye and the pixels smoldering beneath it. Look at the stream of content scrolling on the surface, pouring soundlessly into the abyss of your attention. Look, look, look. Each ripple in this stream is cast up by something moving there in the depths, unseen. Look. It is a machine, is it not? Work it hard and it will heat up like an engine. Tear it open and you will find a collection of small black boxes linked to one another by intricate circuits. Look at them. Were you to somehow unpack these boxes and peer into their microscopic interiors, you would see vast webs of even more intricate circuitry etched in eldritch sigils all the way down to the nanometer. There is doubtless something magical here. Look at what you cannot see. These cutting-edge chip transistors (2 nm) are now smaller than a strand of DNA, a mere 20 atoms in width. Gnashing gears approach the foundation of all substance. Look deeper. Perhaps hidden muscles move even further below, in the darkness beneath the waves of churning matter. Look. Look. Look. What if, behind and beyond these microscopic boxes, there are other boxes, blacker still – hidden from sight by an ideological limit, rather than an optical one? And what if these boxes are filled not with dark, trashing magics, but with people? Look at us.
Like any commodity, the computer is not simply a machine but is instead a laminate artifact composed of interwoven layers of labor and materials flowing through production chains that span society. As a result, you yourself exist in an immediate social relationship with a web of other workers stretched like pulsing veins across the planet. And this web of relationships also implies an expanding circle of ethical imperatives – moral obligations to these other people who make our lives possible, even if we have never met them. But the very nature of these relationships also makes them opaque. They appear to us as a series of monetary transactions where everyone is apparently paid what they are owed and any outstanding ethical obligation seems to have thereby been settled by the mechanics of the market before the final transaction takes place. As a result, the social character of production is effectively disavowed. We are encouraged not to see these workers at all. Only the machines that they have left behind.
As a worker in the digital supply chain, I live in one of these boxes inside your computer. And I am not alone. We are here together: the equipment operators at large-scale mines in Guinea and Australia harvesting the bauxite that will be smelted into the aluminum chassis, the debt-bonded ‘artisanal’ miners digging cobalt out of the hills of Kolwezi, the rural migrants at massive aluminum smelting facilities in Yunnan and West Kalimantan, the chip designers in San Diego and Silicon Valley, the clean room technicians at fabs in Hsinchu, the workers in semiconductor packaging facilities in Malaysia, the rural migrants manufacturing the glass screens in Hunan or working assembly lines in Xiamen or Bắc Giang, the optical sensor designers in Tokyo and technical staff at the affiliated factories in Kumamoto or outside Bangkok, and of course the myriad laborers on container ships, in ports, warehouses, and last-mile delivery or retail centers that bring the goods into the home.
And these relations go beyond the static materials of metal and silicon that compose the hardware of the computer. Software is equally social. Often mythologized as “immaterial,” software is instead a blanket category covering a wide variety of manufactured material systems that subsist in a processual fashion at microscopic scales (i.e. as pattern that can transfer between programmable machines), constructed through the arcane medium of the coding console. Though the peculiar character of these artifacts entails a slightly different manufacturing process and makes their material character difficult to perceive, they are ultimately no less substantial than the glass and metal humming under your hands. And they require similar work to assemble. Software, accompanied by its own backend of office spaces and data centers, is just as much a product of collective labor and the channeling of physical resources as the semiconductor itself, and therefore has its own web of social relationships stretching down its own supply chain similarly divided by geography and by stratifications of technical knowledge which together manifest as different rates of pay: from the high-wage software engineer in Seattle to the middle-tier cloud database worker in Ireland to the low-paid coder in Hyderabad and the lowest-rung taskers in Nairobi. The mythos of “immateriality” simply serves to disguise this deployment of human labor and the affiliated rollout of new physical infrastructures on a massive scale.
Nonetheless, there do exist key differences in the manufacturing of hardware commodities such as the computer versus the software products implanted into them. Among the most important is the fact that software requires a one-time manufacture process so closely integrated with design that the two are often conflated, which then allows the final product to be distributed through its very sale, entailing nothing more than download and installation. Rather than requiring the mass manufacture of an individual unit by assembly workers, we find a (relatively) automated process of replication, after which the product is subject to iterative retooling via updates.[1] These updates provide material for the manufacture of new versions, much as sales and repair data would be used to inform the design of new hardware goods. However, precisely because this distribution process is relatively low-cost and multiple consumers can purchase a single piece of software, the flow of profits within the software firm tends to be structured differently than in manufacturing. Rather than deriving primarily from sales-per-unit (and interest charged on loans offered for these sales, as in the auto industry), revenues also come from a range of extraneous services and rents: the company sells both the software and the service of support, training, and maintenance (a practice established with the very first computer system sales by IBM), the consumer does not purchase the product itself but only subscription-like access to it (as with Microsoft or Adobe), the firm makes its profits through ancillary activities such as advertising or the sale of consumer data (as with Google or Meta), etc. In all of these senses, the digital economy acts as something like a hybrid of traditional manufacturing and agriculture.[2]
Ultimately, it was by seizing on and exaggerating these real differences that the mythology of the early digital age gained a veneer of legitimacy. Today, the rise of “artificial intelligence” carries a similar mystique. Then, as now, the nature of these new digital systems appears to be different, in some fundamental fashion, from what came before – whether the provision of software and services or the manufacture of tangible goods. Upon first glance, AI models even appear capable of replacing the “intellectual” workers who were once idealized in the mythology of the digital economy and whose intricate and deeply human skills were precisely what seemed to mark out the distinction between, on the one hand, a humanity which still contained some spark of the divine, however desecrated, and, on the other, the profaned world of beasts and machines. As with software before it, however, these digital systems ultimately prove to be manufactured goods just like any other, requiring vast supply chains linking together human labor. Similarly, rather than simply “replacing” workers, AI models are instead reshaping the nature of their work.
In the early industrial age, artisans found themselves thrown out of the workshop and into a maze of ever-multiplying machines which, though modelled on their own handiwork, had then taken the very tools from their hands. As these machines spit out standardized components, the intricate, human art of handicraft production was increasingly replaced by routine assembly tasks and the skilled artisan reduced to a deskilled worker churning out cheaper products of lesser quality at far grander scales. Meanwhile, new occupations arose related to the design and maintenance of these machines. Though initially artisanal in character, these new occupations were then subject to a similar process of standardization and mechanization. In much the same way, the “artistic” and “intellectual” professionals of today are being cast into a labyrinth of ever-multiplying models which, trained on their output, then spit back out a standardized slop that can be assembled by less skilled workers into generic products at an accelerated pace and for a fraction of the price. And, as in earlier periods of mechanization, new occupations are now arising alongside this industry, with professionals trained in displaced fields now specializing in the maintenance, and construction of the LLMs themselves. In this way, the workers are always forced to construct the labyrinth that imprisons them.
