Google says that its researchers have created an AI "child" that is able to outperform any human-made counterparts using a reinforced learning scheme. This scheme enabled the AI child to learn a task and then report back to the parent to learn how to better perform. The AI child is very good at recognizing objects including people, cars, handbags, traffic lights, and more in real-time.
The team of researchers points to the AI child created by the supercomputer as proof that some machine-made programs are more accurate than human-created ones. The AI child is dubbed NASNet and is controlled via a neural network called AutoML made by Google Brain. Google Brain is the same outfit that created a neural network that is able to invent its own encryption.
NASNet is 82.7% accurate in predicting images correctly, which is a gain of 1.2% compared to previously published results. These results are looked at as proof that automation could allow the creation of more AI without human input. This is how Skynet begins.
The ability for a supercomputer to build its own AI that is better than what humans build will do nothing to make those who fear AI run amok feel better. Elon Musk has been particularly vocal about the risks runaway AI poses to humanity. Musk has said in the past that AI poses a "fundamental risk to the existence of human civilization." Musk isn’t alone in seeing AI as a massive risk, Stephen Hawking also sees AI as potentially the worst event in human history.