Amazon Fresh has decided to shelve its “Just Walk Out” technology that let customers fill their carts with groceries and skip the cashier, with their accounts charged later. The system relied on an apparatus of cameras and sensors set up all over the stores to track their purchases. An Amazon spokesperson told DFD it will “deliver even more convenience to our customers” with a simpler solution: shopping carts resembling a self-check-out on wheels that requires customers to manually scan items. This closes a chapter of what Amazon scientist Gérard Medioni had explained last year was “magic” powered by a complex array of buzzy technologies: object recognition, computer vision, deep machine learning, and of course, generative artificial intelligence. “Creating that magic—determining who took what—is harder than you think,” he said. Apparently, it was harder to pull off than even Amazon thought. Amazon Fresh’s redesign drew attention to what really went into the “magic,” a question at the heart of an over-a-year-old report from The Information. According to the story, as of mid-2022, Amazon had more than 1,000 workers in India manually reviewing Just Walk Out transactions and labeling video images to train its model. For every 1,000 sales, the technology required about 700 human reviews — much greater than an internal target of fewer than 50 reviews, the report claimed. (Amazon disputes these figures). Quickly, skeptics pounced on Just Walk Out as the latest sign of overblown AI hype, and social media observers questioned it with a sense of betrayal: did what they believe to be automated technology actually run on humans behind the scenes, and is that why their receipts would occasionally come hours later? “Over and over, I see these AI executives underestimate how complicated the real world is,” entrepreneur and AI scientist Gary Marcus told DFD in an interview. “Maybe 20 years from now it'll work, but I think computer vision is not as good as the field is trying to imply.” Amazon says no, workers in India have not been watching shoppers live. A company spokesperson did confirm that humans “validate a small portion of shopping visits by reviewing recorded video clips.” But it insists that doesn’t happen in real-time — only after customers leave the store, and in cases when the computer vision needed human input to accurately determine what people bought. Some AI industry insiders saw this coming. After walking into an Amazon Fresh store in 2022, Emmanuel Maggiori, a machine learning software engineer, predicted that a fully cashierless solution was too good to be true. “There are certain problems that are very, very hard to solve, and this is one of them because of all the edge cases,” he told DFD. “Current AI is based on machine learning, which learns from examples of how to do the job.” Just Walk Out, for example, uses generative AI to simulate shoppers doing different activities and train its model’s response. But Maggiori said the model can’t possibly cover all real-life scenarios without its own awareness of how the world operates — a not-yet-ready development closer to artificial general intelligence: “I don't think it’s a question of time, trying more and more. It’s more: can we find a new methodology that will work better?” For all of its touted advancements, AI has an ongoing dependence on invisible workforces. Marcus drew parallels to other tech services that he said shared Amazon’s “fake-it-til-you-make-it” approach: Cruise — whose driverless rides were found to require frequent assistance from humans working as “remote assistants” — and OpenAI’s ChatGPT — which used outsourced Kenyan laborers earning less than $2 per hour to scrub itself of toxic text. Both Cruise and OpenAI have confirmed their use of human labor. Social media companies, too, depend on a blend of algorithms and human workers to moderate content. Growing evidence suggests that job inflicts a psychological toll on labor that’s often low-paid and outsourced. Christopher Andrews, a sociology professor at Drew University who has researched the adoption of self-checkout, pointed to a lesson from the past. The founder of the original self-service store, Piggly Wiggly, tried in 1945 to stand up another grocery store staffed by robots. But it closed after less than a year due to operational costs and customers finding the experience strange. “There is a lesson there to be learned, specifically that technology does not operate magically and independently of humans,” Andrews said. “Most of the thinking and focus has been on the front end of stores, the checkout lane, and cashiers - not on the costs to operate and maintain the technology or the forms of work it might create.” That underscores a broader critique of these automated retail technologies. Even if they functioned as advertised, it’s not clear if the vision matches what customers want from their grocery stores. In fact, Amazon says that while customers enjoyed skipping the checkout line, they also wanted to keep tabs on how much money they spent and saved as they shopped, instead of getting a receipt after they left the store. It’s planning to keep Just Walk Out in smaller stores and quick-stop settings like stadiums and airports. Ultimately, the promises tied to a 21st Century story may be more for the benefit of corporations than customers, which could be part of the problem, according to Andrews. “[R]etailers cannot shake the dream of finding ways to use technology to reduce labor costs,” he told DFD. “All of this is part of a vast social experiment in which retailers, airlines, banks, and restaurants are trying to see if they can socialize customers into doing some of the work previously performed by paid employees by framing it as ‘faster’ or ‘more convenient.’”
| 
Comments
Post a Comment