Thursday, May 14, 2015

Baidu's Brain & AI's Rise

Although it's incremental, the improving quality of AI is becoming more and more obvious. This Technology Review article is about Chinese search giant Baidu's new super computer, built specifically for massively parallel "deep learning" techniques, advanced neural network methods which are applied to image recognition. On a standard training/recognition benchmark, it improved only slightly on Google's recent accuracy record (4.58% error rate vs. 4.92%, in March). But the average human performance on the data set is 5.1%, which was first bested by Microsoft (4.94%) in February! This is not factoring large numbers or playing chess, it is pattern recognition, the hallmark of human intelligence. 

We are definitely on a slippery slope here.  It's not just SF, though the SF writers remain well ahead of the technology when it comes to extreme extrapolation and scary scenarios. But even in reality, the accelerating rate of change and the number of different entities in the world working on this stuff suggest we may be in for a wild ride over the next few years. Will there be unintended consequences? Probably. 

For a good overview of what AI means and where it seems to be going, I heartily recommend two recent articles by Tim Urban, author of the wonderful website Wait But Why. "The AI Revolution: The Road to Superintelligence" and "The AI Revolution: Our Immortality or Extinction" are well researched, well explained, and frankly a wee bit scary, even to someone who has followed AI development in fact and fiction for decades. These articles caught the attention of Elon Musk, the billionaire founder of SpaceX and Tesla, an optimistic futurist if there ever was one. Musk believes that ASI (artificial superintelligence) represents an existential risk to humanity beyond even nuclear and biological weapons. 

As a nice side effect of Tim and Elon's shared interest in ASI, Tim is now doing a new series of Wait But Why articles on Elon Musk's ideas, technologies, and companies. Tim has become one of my favorite writers, and Wait But Why one of my top 5 web destinations. Great stuff on a wide range of subjects. If I ever wanted to try to turn my little wooden-headed Pinocchio blog into a "real boy," it would probably be a clone of Wait But Why. 


Unknown said...

Pattern recognition may be the hallmark of human intelligence but is also abused -- stereotypes, racial profiling. Hope AI can avoid those outcomes.

FlyingSinger said...

I hope so too. But a higher priority goal will be for artificial super intelligence (ASI) that considers humans to be colleagues and friends and not dim-witted resource wasters. This is by no means a sure thing, especially for second generation ASI's that will designed by first generation ASI's, perhaps just days after the first generation attains superintelligence. The third gen ASI's could come even faster and have even less empathy with their "grandparents" (i.e., we slow-brained bipedal primates). To learn well and be useful, AI's will need to be able to define and redefine their goals. We won't understand them well enough to control this for long. Once their goals start to diverge from ours, no one knows how it will go. We need to design AI's so they remain friendly to us even when they become trillions of times smarter than us. No one knows how to do this either. So we may have a serious problem. An existential threat spreading itself throughout the internet. Have a nice day! :)