Google has been going hard in the artificial intelligence scene recently. Their name boasts such achievements as generating sounds that humans have never heard, creating an AI on a chip that could rival NVIDIA’s, beating the world’s best human Go player, and news with Google’s name beside “AI” seems to be pouring out every week. It’s safe to say they have a lot of expertise (and funding!) in the artificial intelligence world.
It is their continuous effort in this field that has brought us one step closer to the singularity with their most recent announcement–Google’s AIs are now creating their own “child” AIs.
These AI designers of artificial intelligence code their “spawn” on par with the world’s leading experts in machine learning, but do so at a much quicker rate. Basically, the parent AI proposes a child architecture for a particular task then tests it, re-evaluates, and redesigns the code based on what it learned–sometimes thousands of times.
Google’s CEO Sundar Pichai called this “AI inception”–AI using AI to design AI–but the actual process of artificial intelligence designing deep neural nets is referred to as “AutoML” for automated machine learning.
These AutoML processes are not only able to handle what has previously been among the most arduous tasks in designing an effective artificial intelligence, it has brought to light potential solutions that researchers had previously written off as impractical or unsuited for their applications, making AI a truly intelligent and integral part of its own design process.
Another interesting aspect of AI-designed AI is the presence of code snippets that remain in their designs but have no discernible purpose to human coders.
Because the ability of artificial intelligence to design and train other, task-specific AIs is nearly on par top-notch experts in AI, it is apparent that we are reaching an important crossroads–one where AI can be truly available to everyone.
Although AutoML processes on this scale are still expensive and resource-intensive at the moment, the steps that Google is taking now are the first ones on a long path of continuous improvement for AI. Removing the middleman, the expert in AI which is, of course, a rare breed of talent, may let even your average person begin designing artificial intelligences to solve their own, personal needs. When AI gets into the hands of the general public and without the major costs associated today, we may see more widespread adoption and AI assistance on a plethora of tasks we’ve not yet thought about.
Of course, there could be dangers in releasing a self-designing, -improving, and -teaching AI to Joe Blow, depending on his intentions. Ethicists around the globe are constantly working on ways to ensure that AI works alongside us and in harmony with humanity and cannot be turned against us. Even tech moguls such as Elon Musk are designing companies such as Neuralink to keep AI from becoming “other” from mankind.
Despite fears that are prominent in the media regarding artificial intelligence, the technology will continue to grow and become even more commonplace, and perhaps even curiouser.
Artificial intelligence is now being embedded with an artificial form of curiosity that helps them grow and learn even without a particular knowledge goal in mind. Combine these technologies, and we may begin seeing solutions that we could not yet have even dreamed of.