About 542 million years ago, something weird and profound happened on Earth.
Quite abruptly, life went inventively crazy, proliferating from simple, rudimentary single-cell organisms into myriad multicellular forms. Evolution discovered the idea of more sophisticated and specialized cells, and most of the basic body plans we know today. Biologists call it the Cambrian explosion.
Human technology may be on the verge of a similar explosion.
The technological upheavals of the last few centuries – the Industrial Revolution and the information age – have been relatively tame in the sense that people have remained the creative force, even if using different tools and working in different ways. That may finally be about to change, according to Gill Pratt, program director for robotics research at the Defense Advanced Research Projects Agency (DARPA).
He makes a good argument that we’re about to have a Cambrian moment in robotics and artificial intelligence, as devices become able to do anything people can do, including thinking.
To be sure, robotics enthusiasts have been making such predictions since the 1950s. What makes now different is the amazing and totally unexpected progress of research in the past few years. The buzzwords are “deep learning” and “cloud robotics.”
Even five years ago, algorithms for computer vision could barely recognize even simple objects, such as a ball or a square block, especially in realistic environments. Now, such algorithms can easily distinguish different breeds of dogs, and can recognize human faces as well as or better than real people can. Neural networks are breaking image-recognition records on an almost weekly basis.
Last month, Google announced that it could translate text from more than 20 foreign languages just from photos – say, from a menu or handwritten note. Earlier this year, Google’s deep learning group demonstrated that an algorithm could learn to play all of the Atari video games of the 1980s with the same skill as human experts – simply by watching and learning.
The neural networks differ from those explored in the past in several fundamental ways. In particular, they use insights learned from modern neuroscience about how the brain stores and reworks old knowledge. Pratt, who has followed these developments as closely as anyone, expects that robots, following the deep learning ideas, will soon be able to perform “any associative memory problem at human levels.” Any. And that’s just the beginning.
Robots no longer need to learn on their own: They can now share insights over networks, so that powerful knowledge spreads quickly to other robots. Google’s self-driving cars, for example, share maps, images, and data on previous driving experiences and traffic. All the information gets fed to computers in the cloud – that is, distributed over the Internet – where analysis takes place to improve the cars’ performance.
As Pratt notes, robots will soon be able to learn by imagination as well, a trick previously unique to people. Having learned pretty good behavior in a variety of situations, a robot can run simulations to explore circumstances unlike anything it has yet faced. It can experiment with different ways of behaving and find possible solutions. And through the cloud, as he puts it, “every robot’s dreams will improve the performance of all robots.”
Which is why the Cambrian explosion analogy is far from silly. In biology, that explosion happened when evolution somehow stumbled on the pathway of design leading out of the single-celled trap and into new open space, where so much more was possible.
In artificial intelligence, the combination of deep learning methods, advancing neuroscience, computing power and the Internet is doing the same for brains and intelligence, which will no longer be limited to the human kind. What will we find – and what will our super-smart robotic companions find – in this new space?
It’s hard for our limited human brains to imagine how this will change our world. Maybe, as Pratt suggests, robotic capabilities will advance so quickly that most human skills will become redundant, and robots will destroy our economies or create massive inequality. Perhaps, as Nick Bostrom explored last year in his brilliant book “Superintelligence,” robots doing their own research and development, better and faster and smarter than we can, will simply replace us as the new top species.
They’re not only going to be better at number-crunching and mathematics, but at other traditional human skills as well – persuasion, deception and strategy.
Perhaps humans will go the way bacteria did. They existed before the Cambrian explosion, and they still exist today. In fact, they continue to thrive. They just don’t hold the dominant position they once did.
Mark Buchanan, a physicist and Bloomberg View columnist, is the author of the book “Forecast: What Physics, Meteorology and the Natural Sciences Can Teach Us About Economics.”