Structure amid Chaos
Describing Amazon as a big planning machine doesn’t quite match its image as an icon of “new economy” disruption. Even before Silicon Valley became a hub of global capitalism, planning was typically well hidden behind the facade of competition. Today, the facade has only become more ornate: all you see is a website and then a package at your doorstep. Behind the scenes, however, Amazon appears as a chaotic jumble of the most varied items zipping between warehouses, suppliers and end destinations. In truth, Amazon specializes in highly managed chaos. Two of the best examples of this are the “chaotic storage” system Amazon uses in its warehouses and the recommendations system buzzing in the background of its website, telling you which books or garden implements you might be interested in.
Amazon’s recommendations system is the backbone of the company’s rapid success. This system drives those usually helpful (although sometimes comical—“Frequently bought together: baseball bat + black balaclava”) items that pop up in the “Customers who bought this also bought …” section of the website. Recommendations systems solve some of the information problems that have historically been associated with planning. This is a crucial innovation for dreamers of planned economies that also manage to satisfy consumer wants, historically the bane of Stalinist systems. The chaos of individual tastes and opinions is condensed into something useable. A universe of the most disparate ratings and reviews—always partial and often contradictory—can, if parsed right, provide very useful and lucrative information.
Amazon also uses a system it calls “item-to-item collaborative filtering.” The company made a breakthrough when it devised its recommendations algorithm by managing to avoid common pitfalls plaguing other early recommendation engines. Amazon’s system doesn’t look for similarities between people; not only do such systems slow down significantly once millions are profiled, but they report significant overlaps among people whose tastes are actually very different (e.g., hipsters and boomers who buy the same bestsellers). Nor does Amazon group people into “segments”—something that often ends up oversimplifying recommendations by ignoring the complexity of individual tastes. Finally, Amazon’s recommendations are not based on simple similarities, such as, in the case of books, keywords, authors or genres.
Instead, Amazon’s recommendation algorithm finds links between items based on the activity of people. For example, a bicycle repair manual may consistently be bought alongside a particular bike-friendly set of Allen keys, even though the set isn’t marketed as such. The two things may not be very obviously related, but it is enough that some people buy or browse them together. Combining millions of such interactions between people and things, Amazon’s algorithm creates a virtual map of its catalog that adapts very well to new information, even saving precious computing power when compared to the alternatives—clunkier recommendations systems that try to match similar users or find abstract similarities.
Here is how the researchers at IBM’s labs describe Amazon’s recommendations: “When it takes other users’ behavior into account, collaborative filtering uses group knowledge to form a recommendation based on like users.” Filtering is an example of an IT-based rejoinder to one of the criticisms Hayek leveled against his socialist adversaries in the 1930s calculation debate: that only markets can aggregate and put to use the information dispersed throughout society. The era of big data is proving Hayek wrong. Today’s deliberately planned IT systems are starting to create “group knowledge” (collective intelligence, or shared information that only emerges out of the interactions within or between groups of people) out of our individual needs and desires. And Amazon doesn’t just track market transactions. Beyond what you buy, the company collects data on what you browse, the paths you take between items, how long you stay on the page of each item you browse, what you place in your cart only to remove it later, and more.
Hayek could not have envisioned the vast amounts of data that can today be stored and manipulated outside of market interactions (and, to be fair, even many Marxists have assumed that the myriad capricious variables associated with faddish consumer items in particular forecloses the capacity for their socialization), although he certainly would have admired the capitalists such as Bezos who own the data and use it to pad their obscene fortunes. It is a delicious irony that big data, the producer and discoverer of so much new knowledge, could one day facilitate what Hayek thought only markets are capable of.
Really, it is not such a big step from a good recommendations system to Amazon’s patent for “anticipatory shipping.” It has a viability beyond any Silicon Valley, TED Talk–style huckster bombast or tech-press cheerleading. The reason this genuinely incredible, seemingly psychic distribution phenomenon could actually work is not a result of any psychological trickery, subliminal advertising craftiness, or mentalist power of suggestion, but is found in something much more mundane: demand estimation. With its huge data sets that measure the relationships between products and people, Amazon is already very successful in figuring out demand for particular products, down to a previously unimagined level of detail.
The bigger question for egalitarians is whose demand counts, and for how much. Under capitalism, it is one dollar, one vote: those with fatter wallets have a much bigger influence over what society produces, simply through their much greater buying power. We get a few super-yachts instead of superabundant housing for all; and we might well say the same when it comes to which consumer items we prioritize for production and distribution.
In our irrational system, the ultimate purpose of product recommendations is to drive sales and profits for Amazon. Data scientists have found that rather than high numbers of customer-submitted reviews, which have little impact, it is recommendations that boost Amazon’s sales. Recommendations help sell not only less popular niche items—when it’s hard to dig up information, even just a recommendation can be enough to sway us—and bestsellers that constantly pop up when we’re browsing.
Zooming out beyond Amazon’s corporate interests, the recommendations system is a way of managing and integrating great swaths of social labor. Many of us freely, without expectation of any reward, spend time and energy writing reviews and giving out stars to products or even just mindlessly browsing on Amazon and other technology platforms. This is work that we and others benefit from. Even over the course of one day, we may repeatedly engage in unpaid labor to rate everything from the relatively innocuous, such as call quality on Skype, to the more serious, such as posts, comments and links on Facebook and Twitter, to the potentially very impactful on individual lives, such as the “quality” of Uber drivers. Under capitalism, the social labor of many is transformed into profit for the few: the filtering may be “collaborative,” but the interests it serves are competitive and very private.
Workers Lost in the Amazon
While many of us end up using free time to perform the social labor that allows Amazon to perfect its recommendations system, Amazon’s warehouses run on paid labor that is nonunionized and frequently occurs under appalling, similarly big data–disciplined conditions. Before taking a closer view of the work itself, let’s quickly look at the workplace. The focal points of Amazon’s distribution network are its warehouses, which the company calls “fulfillment centers.” These usually take up football fields’ worth of floor space jammed with shelving units. Amazon uses a peculiar form of organization called “chaotic storage,” in which goods are not actually organized: there is no section for books or subsection for mystery fiction. Everything is jumbled together. You can find a children’s book sharing a bin or shelf with a sex toy, caviar next to dog kibble.
Once again, powerful planning is what allows Amazon to save on what turns out to be needless warehouse organization. Every item that enters a fulfillment center gets a unique barcode. Once inside the warehouse, items go in bins, each of which also has a unique code. Amazon’s software tracks both the items and the bins as they move through the warehouse. The software alway
s knows which bin an item is in and where that bin is. Because items can always be found easily, deliveries from suppliers can be unloaded where it is convenient, rather than methodically organized and reorganized.
Amazon’s chaotic storage could be a metaphor for the free market system: at first glance, it seems that the chaos organizes itself. Orders and packages zoom through the system and customers get what they want. But as with the free market, upon closer inspection we see thickets of deliberative planning at every step. Highly refined IT systems make sense of the chaotic storage, track items from the moment they arrive at a warehouse to the moment they leave, and make sure everything falls seemingly supernaturally into place. Everything ordered, coordinated, planned and not a market in sight to perform any of these billions of allocation decisions.
Planning is also present in the most minute details of a warehouse worker’s day. Handheld scanning devices tell workers where to go to pick items for orders. Workers are appendages of machines that lay out precisely which routes to follow between shelves and how long they should take. Here’s how a BBC undercover worker-reporter described the work: “We are machines, we are robots, we plug our scanner in, we’re holding it, but we might as well be plugging it into ourselves.” A leading UK researcher on workplace stress contacted by the same BBC investigation claimed that conditions at Amazon warehouses pose serious physical and mental health risks.
Around the start of this decade, Amazon’s top operations managers determined that its warehouses were still too inefficient, and so they themselves went shopping for something better. In 2012 Amazon bought Kiva Systems, a robotics firm, and it now uses robots to put its entire shelving system into motion. Amazon’s updated, even more automated fulfillment centers now feature shelves that move and humans who stand in place—the opposite of what a warehouse normally looks like. Flat, Roomba-like robots rove the warehouse floor along designated pathways. They can lift entire shelving units just off the ground and maneuver them along the same pathways to “picking stations.” These are small designated areas where human order pickers stand, taking items from storage bins and putting them into order bins as shelving units come and go.
The social, physical and mental cost of a machine for delivering the right things to the right people ultimately falls on the workers who make the machine hum—regardless of whether workers are piloted around a maze of shelves by a handheld scanner or pick orders in place while robots whiz to and fro toward them. The boosters at Wired magazine are in awe of the subjugation of the Chaplins in this twenty-first-century Modern Times: “The packing stations are a whirl of activity where algorithms test human endurance.”
Other more critical reporting has been less kind to Amazon in fleshing out just what these endurance tests entail. In 2011 the Lehigh Valley, Pennsylvania, local paper, the Morning Call, investigated its nearby Amazon fulfillment center. Workers said they routinely faced impossible-to-meet targets, debilitating heat and constant threats of being fired. On the hottest days of the year, Amazon had paramedics on hand outside the warehouse to treat heat-exhausted workers—a cheap Band-Aid solution for Amazon that makes clear its low estimation of health and safety; apparently humane working conditions are not one of its algorithms’ optimization constraints. It was only after this story blew up in the national media and the revelation hurt its largely liberal-tech-and-innovation brand image that Amazon began to refurbish some warehouses with air conditioning. In fact, only one out of the twenty workers featured in the Morning Call story said Amazon was a good place to work.
Amazon workers interviewed by the media consistently report feeling the constant stress of surveillance. Being too slow to pick or pack an item, or even taking a bathroom break that is too long, results in demerit points. Amassing enough of these points can lead to being fired. Soon, this feeling of constant surveillance could become far more visceral: in February of 2018, Amazon patented a wristband that monitors a warehouse worker’s every hand movement in real time. And Amazon pits workers not only against the clock, but also against one another. Warehouses are staffed by a mix of temporary workers hired by subcontractors and permanent workers hired by Amazon. Permanent positions are few, but they come with some security, slightly higher pay and limited benefits; they are dangled as carrots before temporary workers, encouraging competition and overwork, further fostering a climate of uncertainty and fear.
With the help of robots, the average time to fill an order in a warehouse automated by Kiva technology has plummeted from ninety minutes down to fifteen. Working conditions, however, haven’t budged: the work remains as dull and draining as ever, warehouses remain hot, and the pace of work can be absurdly fast, regardless of the level of automation. While workers in automated warehouses stand all day and try to keep up with the robots zooming by, workers in the nonautomated warehouses can expect to walk nearly double the distance on a daily shift of a typical mail carrier. Even small things like distances to break rooms can be an obstacle—sometimes so long that going both ways can take up most of a break.
Long hours for low pay are the norm in an Amazon warehouse, but the relatively highly paid white-collar workers at Amazon also face a crushing work environment. A 2015 New York Times exposé revealed an environment of overwork and “purposeful Darwinism” that pushes many past their physical and emotional limits. Even if sophisticated planning is Amazon’s workhorse, it is implemented within the bounds of a ruling ideology of ruthless competition that breaks white-and blue-collar workers in different ways. Put differently, Amazon is doing exactly what Marx described in a lesser-known passage from The Communist Manifesto: “The bourgeoisie cannot exist without constantly revolutionizing the instruments of production, and thereby the relations of production, and with them the whole relations of society.” Our task must be to disentangle the good brought by technology from the tentacles of a system that degrades workers and subverts more rational planning.
Amazonian Technologies beyond Amazon
Despite being a model of the new, disruptive, internet-dependent capitalism, Amazon remains a planning device as much as other companies ever have. In simplest terms, Amazon is a giant planned machine for distributing goods. It is a mechanism for forecasting, managing and meeting demand for an incredibly wide array of things we need and want. It is a collection of thousands of interlocking optimization systems that work together to carry out the deceptively simple task of moving objects from producers to consumers. Rather than the anarchy of the market, once we enter the Amazon, we are entering a sophisticated planning device—one that offers not only clues for how we could manage demand and supply of consumer goods in a society not built on profit, but also warnings to would-be planners for the public good.
British economic journalist Paul Mason suggests as much in his 2015 book, PostCapitalism, imagining a future where the data accumulated by Amazon and other large consumer-facing firms is used to regulate production. His vision is one where comprehensive planning takes the place of separate and haphazard supply and demand. For Mason, capitalist technology will eventually be the means that allows us to go beyond the system that created them. Socialist construction, however, is not so simple. Instead of optimizing the satisfaction of our needs and desires, as well as workers’ working conditions and livelihoods, Amazon’s plans are geared toward maximizing profit for its shareholders—or future profit, since Amazon keeps plowing money from sales into research, IT and physical infrastructure to squeeze out competitors. Planning for profit is in fact an example of capitalism’s web of allocation inefficiencies. The planning technologies dreamed up by Amazon’s engineers are a way of meeting a skewed set of social needs—one that ends up enriching a few, misusing substantial free social labor, and degrading workers. A democratized economy for the benefit of all will also need institutions that learn about people’s interests and desires, optimize via IT systems, and plan complex distribution networks; but they will look different, perhaps alien to the systems we have today, and they will strive towar
d dissimilar goals.
Three challenges should give us pause before even beginning to call the riddle of democratic planning solved.
First, there is large-scale technical feasibility. The difficulty of planning and optimizing even the isolated task of delivering Amazon’s packages demonstrates that designing systems for economy-wide planning will be anything but trivial. The algorithms that power everything from Amazon’s recommendation system to Google’s search engine are still in their infancy—they are relatively simplistic, making best-estimate guesses, and are prone to failure. Algorithms run into systemic problems, for example with working class and poor people who more frequently use shared devices to shop or non-English speakers, where their capacity for “reading” nuance is limited. We’ll have to storm both the barricades and the optimization problems.
Second, the planning done by Amazon and others still relies heavily on prices in interactions that take place beyond the borders of the firm itself. Amazon purchases its inputs—from the multitude of items it stocks, to the warehouse shelves they sit on, to the servers that run its database—on a market; consumers, meanwhile, also take into account the relative costs of items when deciding whether to add them to their virtual carts. Beyond the confines of the firm, a market system continues to operate. This means that it’s not simply a matter of repurposing existing technologies, lopping off the bosses and otherwise keeping everything the same.
Even though there is market-less planning within corporations, it is a form of hierarchical, undemocratic planning that is very much necessary to survive and thrive in a market. Many elements of this planning apparatus, their very form and purpose, are conditioned by that undemocratic hierarchy. A democratic planning system built from the ground up would look very different. To catch a possible glimpse, even foreshadowing, of what a market-less world might look like, compare Amazon’s book section to an online public library catalog. A library catalog also contains a vast, searchable, interconnected array of books—but not a single price. And it should be possible to harness far more information than is currently contained in a library catalog: for instance, how long people spend looking at a book, (with digital books) how many of its pages they actually read, whether they click to see if it is available in their neighborhood, whether they are willing to place a hold (and, for instance, to do so even if there are ten others in the queue in front of them) and what path they follow through the online catalog. The example of an expanded library catalog shows that we could build not only recommendation tools, but also models of interests, demands and needs that are independent of prices.
The People's Republic of Walmart Page 8