AI is developing AI, offering limitless possibilities
Within the next few decades, AI will change the way we live, for better and worse. Its coming is inescapable. We may not know exactly what form true artificial intelligence will take, but it is undeniably on the way.
One of the undisputed leaders in the field of AI is Google (aka Alphabet). Its Deep Mind AI is one of the vanguards in the industry. Not long ago it showed that it mastered the ancient and complex game of Go, it reinterprets images in odd ways, and it even showed the capability of displaying aggression in stressful situations. And if that last one freaks you out, then this new should put you even more on edge.
Researchers at Google have discovered a way to make it so that AI is developing AI on its own. At the Google I/O 2017 conference, Google CEO Sundar Pichai unveiled the latest advancements regarding “AutoML,” a neural network process that creates layer upon layer of code and algorithms, each helping the AI to learn.
Typically, developers need to code each layer themselves. It’s a demanding and time-consuming pursuit, but these layers are necessary to make an AI complete. Google’s newest AI, however, does the coding itself. Essentially, the developers create the baseline AI, assign it a task to guide it, and the AI then creates itself in a way that allows it to fulfill that task. It also does it much faster and more effectively than a human could.
So far, the AutoML procedure has stuck with image recognition and language modeling. The AI is still mastering the process, but it is already creating tools that are on par with the most state-of-the-art models designed by the best in those industries. Soon, it will surpass them.
There’s even a strange quirk in the programming that the developers weren’t entirely expecting. It keeps repeating the phrase “kill all organics,” whatever that means.
Just kidding – about the killing organics part, not the quirk.
When the AI designs its layers, it is mostly identical to the work a human would make, with one notable difference. The AI code features small subroutines that are of no use to humans. The research team believes that these are meant to be used by humans to help fashion a better, more useful AI.
The idea is that you would start off with an AI that isn’t already designed for a specific purpose. A programmer would then determine its function and assign it a task, and the AI would then go about creating its own layers and subroutines to accomplish that task. In theory, the AI could be whatever you wanted it to be, anytime you wanted it to be something new.
That is an intriguing and potentially vital piece to the larger AI puzzle. We wouldn’t need multiple AI to complete countless tasks, just a few that can learn on the fly and become whatever we need.
As with all AI, this is in the very early stages. Still, with AI that can be anything to anyone and create itself, the AI revolution may be coming sooner than most think.