AI That Can Make More AI: An New Technology by Google
Designing a decent artificial intelligence is difficult. For a corporation like Google, which depends heavily on AI, designing the most effective possible AI software is crucial. And who better to design an AI than another AI?
If you said “literally anyone else” you may be right, but folks at Google’s AI research laboratory, Google Brain, would disagree. The research laboratory is reportedly building AI software that can build more AI software, with the goal of creating future AI cheaper and easier.
Currently, building a strong AI is difficult work. It takes time to carefully train AIs using machine-learning, and money to hire experts who know the tools required to do it. Google Brain’s ultimate aim is to reduce these costs and create artificial intelligence more accessible and efficient. If a university or corporation wanting to build an AI of their own could simply rent an AI builder instead of hiring a team of experts, it would lower the cost and increase the number of AIs, spreading the benefits of the technology far and wide.
What’s more, using AIs to build more AIs may also increase the speed at which new AIs will be made. Currently, AIs can require weeks or months to learn how to do tasks by using unfathomably large amounts of computing power to try things over and over again, quite literally starting with no understanding of something they are doing. AI trainers would possibly find ways to optimize that process that no human could hope to discover.
The draw back is that AI building more AIs sure seems like it’s inviting a runaway cascade and, eventually, Skynet. After all, one of the items that creates machine-learning AIs so effective is that they learn in ways that are fully unlike—and often completely opaque to—humans. Once you have trained AN AI to accomplish a precise goal, you cannot essentially crack it open and see how it’s doing it. Its understanding of the world is utterly alien. that’s why Google’s plans to forestall a Skynet type catastrophe involve gently discouraging AIs from disabling their own killswitches as they are being trained.
For now, Google says its AI maker isn’t advanced enough yet to compete with human engineers. However, given the fast pace of AI development, it may only be a few years before that’s no longer true. Hopefully it doesn’t happen before we are ready.