Machine learning is tough — it combines high-level math with complex conceptual thinking, requires both a super-sharp analytical mind and the ability to think widely and creatively. The field is so difficult, in fact, that it seems to be subverting some of the very most basic principles of capitalism itself; no matter how competitive the salaries on offer, universities just don’t seem to be able to pump out enough experts to provide all the expertise modern corporations say they require to advance their services. Now, Microsoft has formed the Machine Teaching Group, an innovative attempt to solve this problem by making it easier than ever to contribute to machine learning. The question is, should we even want them to succeed?
Over the past few years, there have been quite a few doom-and-gloom projections for technological unemployment. The most famous came from Brynjolfsson and McAffee, arguing that a majority of modern professions might soon be able to be computerized via algorithms, robots, or both. Machine learning is the process by which computers can actually start to fulfill those predictions, acquiring new, unthinking abilities through unbiased sifting through the results of experience.However, as mentioned, programming new and better machine learning algorithms takes more than a mathematical ability to do so — you also need some abstract understanding of the best way to approach all the various fields of human knowledge and skill. That would logically lead to professionals from these fields contributing to machine learning algorithms, but then the reverse problem applies and the high programming and mathematical barrier stops the creatives from truly helping out.
The Machine Teaching Group is an attempt to fix this conundrum, lowering the barrier between non-programmers and contribution to machine learning. Microsoft refers to this as the perfectly palatable process of “democratization,” but it also has some darker implications: if Microsoft is as successful as it wants to be with this project, it could push forward the threshold of “computerizable,” encompassing even more jobs than Brynjolfsson/McAffe and colleagues predict. Their models generally assume that the more creative and non-linear a job process, the harder it will be to computerize — the Machine Teaching Group aims to erode even that small comfort.In some fields, like diagnostic medicine and surgery, it will be an unequivocal good, and any labor hardships that come about will be roughly meaningless next to gains in public health. For other professions, though, the calculus is not so clear. To me, the image of a chef contributing to an algorithm on cooking is not unlike the image of a supermarket clerk stocking bags at an automated checkout kiosk. You can’t blame them, or stop them — just hope that the inevitable process they’re a part of somehow works out well for them, and for you, in the end.
Like the mechanical looms of old, progress in this area could help just as many people as it hurts. It won’t necessarily create as many jobs as it destroys — a fallacy about the job market born before the onset of true computerization — but it could allow real decreases in the cost of living, and the proportion of a person’s life that must go toward work. Combined with 3D printing and self-driving cars, robotics and machine learning stand to let companies lay off huge fractions of their workforce. If the system still works to any extent, that should let prices drop dramatically, and offset the poverty it creates by lowering the overall cost of living for everybody.
That may be a utopian dream, but that basic prediction — that gains made during the coming workopolypse will trickle down to benefit the regular Joe — underlies every attempt to replace highly salable skills with programs and technology. If it sounds naive, then we’ve got a problem; when a company the size of Microsoft sets itself to widening the existing bottlenecks in global computerization efforts, when it starts trying to make this sort of middleware solution for the ongoing destruction of work, our projections for possible impacts have to get a whole lot of more proximate.
