When you’re a big acquisition-hungry corporation like Google, sometimes you make mistakes — you pay billions for bug-riddled Nest technology, for example. But really, you’ve got accept such losses as inevitable when you’re pursuing the big, infrequent pay-offs that only modern technology can provide. Increasingly, it’s becoming clear that the $500 million acquisition of DeepMind, which signaled to many observers the true beginning of AI as a major technology industry, is one such winning investment. Not only can Google rent out the company’s services for enormous profit, competing with other major machine learning entrants like Amazon for a quickly growing market share, but it can also use DeepMind’s insights to improve its own competitive advantage.
This week, Google announced that DeepMind researchers have allowed them to take an incredible step forward in the energy efficiency of their data centers. DeepMind was able to analyze their performance and recommend a number of improvements in the placement and use of things like windows and cooling units. The practice saved Google “percentage points” on their use of power — and since Google’s data centers use multiple gigawatt-hours of power each year, that’s nothing to sneeze at.
But the larger point made through this announcement is about the ability of modern artificial intelligence to adapt to just about any problem, and improve on the best mankind has managed on its own. Cooling efficiency in data centers is not some arcane, poorly-studied aspect of engineering; it’s something people get paid a lot of money to think about. And here comes DeepMind, fresh off of beating the world’s best at Go just to make a point, improving on these expert designs.
So, DeepMind’s basic approach can be quickly and effectively adapted to the temperature of air currents around a data center, and the flow of heat through air vents — what else could it improve? The flow of air through physical space is far better understood than, say, the flow of shipments through major ports, or of bee populations across a continent, or of students through university programs. We should expect that assigning a sufficiently insightful AI to find and eliminate inefficiencies in cargo routing, or field-pollinating, or graduate-producing, will produce more than a few percentage points of progress. The need for warehouse managers, public agricultural engineers, and academic advisers is about to plummet right alongside the more traditionally threatened sectors like customer service and manufacturing.
These are the sorts of wide-ranging improvements that could allow continued increases in worker efficiency that match those we’ve seen over the past several decades. But the historical prediction of science fiction — that employment will be calibrated to a certain amount of work done, and thus that increasing the work possible per hour will reduce the overall number of hours in the job — has not generally come to pass.
George Jetson once famously complained to his boss that he’d worked for two hours the day before. What does Spacely think he’s running, asks George’s wife Jane: “A sweat shop!?”
It’s hard to imagine such optimism about the plight of the middle class worker, these days, despite the fact that the technological trends of today provide a much stronger basis to forecast the end of (hard) work than those of the 1960s. The momentous social changes occurring back then were so obvious that even children’s television had to address them, but these days the system is arguably suffering even more violent turmoil that.
It’s not just the endless process of downsizing more and more unnecessary workers, but a growing realization that pure tech solutions can actually be more palatable than human ones. Facebook recently learned just how badly people will react when they learn a human being is manipulating their thoughts — the backlash would likely have been minimal if the revelation had been a buggy or poorly designed algorithm that Facebook simply vows to improve. Google only gets away with reading our email to target ads because it’s only a robot that reads those emails, not a human being. Increasingly, there are very real incentives for companies to switch to a tech solution that has nothing to do with efficiency or cost reduction.
The fruits of applying truly advanced machine learning, from fully reactive AI voice interfaces to decreased power bills, will change the corporate world in a big, noticeable way. The only question is, how?