Google and DeepMind created an AI program to suggest energy saving measures for their data center cooling operations back in 2016, and now the two are handing over nearly complete control to the machine. The way it works is much the same as it has been since the program was created back in 2016, but with a key difference; instead of sending suggestions to a human operator to be implemented, the AI will now put recommended actions into place on its own. The appointed expert watches over operations to ensure that the AI isn't making any bad decisions, and can step in at any time.
The first step in the process is a mass snapshot of the current state of operations that happens every five minutes. Next, that information goes over to DeepMind's deep neural networks for analysis, where every possible course of action is evaluated for energy savings, efficiency, and of course, safety. Once the AI has a safe course of action, it's verified twice over in the system, then makes its way to the data center as a set of instructions. Actions are checked for safety, efficiency and confidence; if the AI isn't certain that an action will bear the intended effects, it's not implemented and a new one is chosen. Meanwhile, there's an automatic failover system and a human watching in case the AI violates Google's stringent safety rules or fails to make a timely or proper decision.
The automation came about not just as another result of progress in the AI world, but also as a matter of solving an employee pain point. Previously, Google had humans taking AI suggestions for data center cooling on board, and many of the workers involved complained that the AI's wide-reaching changes and new practices became too much to keep up with as it learned and grew. The AI hasn't been quite ready to take the reins until now. With this development, the program will only learn and grow more. Its performance as of this writing makes the data center cooling operations close to 35% more efficient than they were before. That efficiency growth plateaued for a couple of months, but may go on the upswing again in the near future. Should this project continue to operate without any troubles, Google may end up implementing full or near-full AI control in other operations in the near future. AI could also make its way into more of the company's user-facing products, as seen in Android 9 Pie.