Monday, February 12, 2018

Four Examples from the Automation Frontier

Cotton pickers. Shelf-scanners at Walmart. Quality control at building sites. Radiologists. These are just four examples of jobs that are being transformed and even sometime eliminated by the newest wave of automated and programmable machinery. Here are four short stories from various sources, which of course represent a much broader transformation happening across the global economy.

Virginia Postrel discusses "Lessons From a Slow-Motion Robot Takeover: Cotton harvesting is now dominated by machines. But it took decades to happen" (Bloomberg View, February 9, 2018). She describes a "state-of-the-art John Deere cotton stripper." It costs $700,000, and harvests 100-120 acres each day. As it rolls across the field, "every few minutes a plastic-wrapped cylinder eight feet across plops out the back, holding as much as 5,000 pounds of cotton ready for the gin." Compared to the old times some decades back of cotton-picking by hand, the machine replaces perhaps 1,000 workers.

One main lesson, Postrel emphasizes, is that big technological changes take time, in part because they often depend on a group of complementary innovations becoming available. In this case: "Gins had to install dryers, for instance, because machine-harvested cotton retained more moisture. Farmers needed chemical defoliants to apply before harvesting so that their bales wouldn’t be contaminated with leaf trash. Breeders had to develop shorter plants with bolls that emerged at the same time, allowing a single pass through the fields." Previous farm innovations often took decades to diffuse, too: as I've mentioned before on this website, that was the pattern for previous farm breakthroughs like the McCormick reaper and the tractor.

The high productivity of the modern cotton-stripper clearly costs jobs, but although it's easy for me to say, these were jobs that the US is better off without. Cotton-picking by hand was part of a social system built on generations of low-paid, predominantly black workers. And inexpensive clothing, made possible by cotton harvested more efficiently, is important for the budgets of low-income families.

Another example mentioned by Postrel is the case of robots at Walmart that autonomously roam the aisles, "identifying when items are out of stock, locating incorrect prices, and detecting wrong or missing labels." Erin Winick tells the story in "Walmart’s new robots are loved by staff—and ignored by customers" Bossa Nova is creating robotic coworkers for the retail world" (MIT Technology Review, January 31, 2018).

Again, these robots take jobs that a person could be doing. But the article notes that the robots are quite popular among the Walmart staff, who name the robots, make sure the robots are wearing their official Walmart nametags, and introduce the robots to customers. From the employee point of view, the robots are taking over the dull and menial task of scanning shelves--and the employees are glad to hand over that task. Apparently some shoppers are curious about the robots, and ask, but lots of other shoppers just ignore them and navigate around them.


An even more high-tech example is technology which uses lidar-equipped robots to do quality control on construction sites. Even Ackerman explaines in "AI Startup Using Robots and Lidar to Boost Productivity on Construction Sites Doxel's lidar-equipped robots help track construction projects and catch mistakes as they happen" (IEEE Spectrum, January 24, 2018).

On big construction projects, the tradition has been that at the end of the workday, someone walks around and checks how everything is going. The person carries a clipboard and a tape measure, and spot-checks key measurements. This technology sends in a robot at the end of the day, instead, programmed to crawl all around the building site. It's equipped with lidar, which stands for "Light Detection and Ranging," which essentially means using lasers to measure distances. It can check exactly what has been installed, and that it is installed in precisely the right place. Perhaps the area that is going to be the top of a staircase is not precisely aligned with the bottom? The robot will know. Any needed changes or corrections can thus happen much sooner, rather than waiting until a problem becomes apparent later in the building process.

As Ackerman writes: "[I]t may or may not surprise you to learn that 98 percent of large construction projects are delivered (on average) 80 percent over budget and 20 months behind schedule. According to people who know more about these sorts of things than I do, productivity in the construction industry hasn’t improved significantly in 80 years." In a pilot study on one site,this technology raised labor productivity by 38%--because workers could fix little problems now, rather than bigger problems later.

But let's be honest: At least in the immediate short-run, this technology reduces the need for employment, too, because fewer workers would be needed to fix problems on a given site. Of course, ripping out previous work and reinstalling it again, perhaps more than once, isn't the most rewarding job, either. And the ultimate result is not just a building that is constructed more efficiently, but a building that is likely to be longer-lasting and perhaps safer, too.

A large proportion of hospital patients have some kind of imaging scan: X-ray, MRI, CAT, and so on. Diagnostic radiologists are the humans who look at those scans and interpret them. Could most of their work be turned over to computers, with perhaps a few humans in reserve for the tough cases?

Hugh Harvey offers a perspective in "Why AI, Will Not Replace Radiologists," (Medium: Towards Data Science, January 24, 2018).  As Harvey notes: " In late 2016 Prof Geoffrey Hinton, the godfather of neural networks, said that it’s “quite obvious that we should stop training radiologists." In constrast, Harvey offers arguments as to "why diagnostic radiologists are safe (as long as they transform alongside technology)." The parenthetical comment seems especially important to me. Technology is especially good at taking over routine tasks, and the challenge for humans is to work with that technology while doing the nonroutine. For example, even if the machines can do a first sort-through of images, many patients will continue to want a human to decide what scans should be done, and with whom the results can be discussed. For legal reasons alone, no institution is likely to hand over life-and-death personal decisions to an AI program completely.

In addition, Harvey points out that as AI makes it much cheaper to do diagnostic scans, a likely result is that scanning technologies will be used much more often, and will be more informative and effective. Harvey's vision is that radiologists of the future " will be increasingly freed from the mundane tasks of the past, and lavished with gorgeous pre-filled reports to verify, and funky analytics tools on which to pour over oceans of fascinating ‘radiomic’ data."

The effects of technology will vary in important ways across jobs, and I won't twist myself into knots trying to draw out common lessons across these four examples. I will say that embracing these four technologies, and many more, is the only route to long-term economic prosperity.