Sunday, November 12, 2017

Not Stupid - Scary

This is not a windsurfing post. Contrary to what some of you may have thought after reading the title, it's not a political post, either. This post is about something that scares me - and scares me quite a bit, more than the front loop in windsurfing.

I develop software for a living, and a lot of my work deals with complex algorithms. At times, that includes machine learning and artificial intelligence methods like neural networks. A few years ago, it seemed that computers were stupid and would remain stupid.

That has changed. Now:
  • Computers can act with more intelligence than any human
  • Computers can learn such intelligence-related skills hundred of times faster than humans, without any supervision - computers can learn skills that take humans decades to master in weeks.
These statements are based on recent developments in computers playing games - specifically, the game of Go. Go is a board game that is very popular in East Asia, where it is played by more than 40 million people. This includes more than one thousand professional Go players in Japan, China, North Korea, and Taiwan. They compete in a number of tournaments where the winner's purse can be as high as $500,000 (compared to the total prize money of $140,000 at the largest windsurfing event, the PWA World Cup in Sylt).

Compare to chess, Go has much simpler rules. But while computers have been able to beat chess champions since the 1990s, Go has been a much harder problem, partly do to the large number of possible moves that make a "brute force" approach to finding the best move impossible. It took until October 2015 until a computer beat a professional go player in an even match.  A few month later, the next version of the computer program beat an 18-times Go world champion.

This was an impressive feat, but the story does not end here - it gets better. The version of the software was able to run on a single computer rather than a network of computers that previous versions required; it beat professional go players 60:0.

Then came the really crazy improvement: AlphaGo Zero. Whereas pevious versions had been trained with thousands of Go-games played by amateur and professional players, AlphaGo Zero only knew the rules. Over a few days, it played a few million games against itself, and used the outcome of the games for "unsupervised learning". After 40 days, AlphaGo Zero played against the older version that had beaten the world champion.  AlphaGo Zero won 100 out of 100 games!

So - a computer program taught itself in 40 days to reach a level that takes the best human players decades to achieve! That's absolutely amazing.

It's also very scary. If a computer can teach itself to surpass any human at a very difficult mental task within weeks, then "artificial intelligences" that are generally more intelligent than humans suddenly don't look like science fiction anymore. Some of the most intelligent people on this planet, including Steven Hawking and Elon Musk, have warned about the potential dangers - perhaps it would make sense trying to understand what they are concerned about?

I won't delve into that now, but let me give you a few things to think about. The AlphaGo software was developed by Google, and is running on hardware designed by Google. One computer with 4 "TPUs" can beat the best human Go player; in total, Google uses about 2.5 million servers at it's gigantic data centers. Plenty of computing power to learn other things. How about learning about the ethics of one species exterminating tens of thousands of species?

Of course, we don't really have to worry about computers - they can't harm us because we can just turn them off, right? Only if the computers were somehow connected to weapons would be have to worry about those science-fiction scenarios. There may be some military drones around, but they are usually flown by human operators; even if capable of autonomous flight, any firing decisions usually require a human. According to Wikipedia, the "U.S. Military is investing heavily in research and development towards testing and deploying increasingly automated systems". But thankfully, these are still to be controlled by human beings, thoughtfully supervised by the Commander in Chief. Nothing to worry about!
--
If you're interested in learning more about the underlying AI or want to watch the Go games between AlphaGo and Go professionals, check out the DeepMind channel on YouTube.