AI researchers have come up with a cunning plan. They realise that humans are providing the training data and technology and that humans are flawed. So, they are letting AI build both the problem and the answer.
It is, according to one of the founders of this new idea at Open AI, the same as the Darwinian process for machines.
Letting AI use itself to learn is revolutionary. It could be the route to achieving Artificial General Intelligence (AGI), where solutions for one problem are used to solve other problems, so far an unattainable goal.
It is a trigger point for, frankly, who knows what.
Already, AI has solved problems – or won games such as Go – using methods and ideas that humans had not even conceived.
And that is both exciting and frightening at the same time. If it is unleashed to create the most effective solutions to problems unsullied by human frailty, we are very likely not to like the answers.
Essentially we will be governed by machines.
AI is not new, and it is easy to wonder what on Earth we were thinking when we thought that intelligent machines were a good idea. But the genie is out of its bottle, and it is not going back in.
Meanwhile, humans are funding every effort to create AGI. As Richard Windsor reports, investments are being made that look more like philanthropy than real investments. The impetus is huge and getting bigger.
Once AGI is achieved, and we are awed (then enslaved) by the step changes in progress in almost every area where we struggle today, it will be dominant.
Allow properly intelligent AI to solve, say, climate change, and you don’t need AI to see where the problem lies. Humans are that problem.
So what would a machine deem to be the solution?
AGI may still be a long way away, but unleashing the potential of building AI using the technology it designs could accelerate the process a hundredfold. As one of the researchers who leave the program running overnight said, ‘he has no idea what he will find in the morning.’