A Weird Month for Artificial Intelligence

Some of the more interesting and universally accessible developments in technology over the past decade (to me anyway) have been within the artificial intelligence world. Unlike many other industry advancements, stories of high-profile AI successes or failures make suitable discussion fodder at family dinner tables. The "Fork Me On GitHub" types appreciate the complexity of the required engineering efforts, while the "Hey Person Who Computers — Can You Set My VCR Clock?" crowd can just think "whoa" without concern for what is happening under the hood. 

One somewhat recent and well-publicized AI event was the Jeopardy-playing Watson supercomputer built by IBM. Those in the technology community understood the difficulties of natural language processing (all while being required to answer in the form of a question) while much of the world just wanted to see a computer built by anonymous geeks humiliate more smug geeks on live TV. 

Every high-profile AI event comes replete with two varieties of interested observers. There is clearly a vast contingent of technology fans watching keenly while "rooting for the machines", perhaps trying to predict additional uses (and misuses) for AI, while there is at least some population hoping to see a disaster of NASCAR pileup proportions. 

No matter which camp you subscribe to, March 2016 was a good AI month for us. 

The Good

Inspired by Deep Blue's victory over Chess champion Gary Kasparov almost 20 years ago, AlphaGo is a Google DeepMind software project designed to master the ancient game of Go. Unlike most card games (and my wife's mysterious "Laundry Sorting Protocol"), Go has very simple rules yet features "more possible positions... than atoms in the universe". After skunking European champion Fan Hui in the fall, AlphaGo played the world's top Go player Lee Sodol in a best-of-five series this month.

The several hour matches of computer vs human were streamed live from Korea with commentary provided by Go experts in multiple languages. Thousands of nerds, AI lovers, and board game enthusiasts (many falling in all three categories) across the globe were glued to quad monitor setups, which surely created an as yet unmeasured negative impact on software developer productivity much like March Madness has on companies that employ people who watch sports. AlphaGo won four of the five matches, and although there were a few stumbles the project was deemed a major success for AI (and a huge failure for humans).

The Bad AND the Ugly

The NASCAR wreck that some AI observers hope for came ten days after AlphaGo's victory in the form of a chatbot named Tay. Described by Microsoft as "A.I. fam from the internet that's got zero chill!" (whatever that means), Tay was designed to emulate a teenage girl and released to Twitter to interact with the general public. A spunky teen girl chatbot on Twitter - what could possibly go wrong?

Media coverage of the event by the tech press was optimistic early on. Less than 24 hours later, Tay had to be taken down (Tay's tweets are now "protected") due to some rather unfortunate tweets. These included commentary on Nazis and Hitler (Godwin's Law again confirmed), Caitlyn Jenner, feminism, President Obama, and at least one more mention of a Donald Trump policy proposal* than you might find in a Republican debate.

Much like my three-year-old niece, it turns out that Tay was built with a distinct vulnerability to blindly obeying "repeat after me" commands, which led to thousands of the planet's best and brightest testing the exploit. As an aside, I must confess to owning an old Atari system (thanks Google, it was the 1040 ST) in the 80s with a voice synthesizer that had some pretty nasty things to say about my sister. However, some media coverage noted that (oddly) not all of the offensive tweets were prompted by "repeat" commands, including a rather benign tweet about comedian Ricky Gervais prompting a (you guessed it) Hitler comparison.

What Did We Learn?

AlphaGo's success has proven that there are probably few traditional games where humans will be proven superior to machines. I heard some unsubstantiated rumors on LinkedIn that Google is currently working on a top secret AI project capable of defeating humans in one of the more complex card games of the past century — Cards Against Humanity. Don't expect AlphaGo to be invited to anyone's game night, at least until it can drink a beer.

Microsoft pitched Tay with the tagline "the more you chat with Tay the smarter she gets". It turns out that wasn't true.

* Tay (@Tayandyou) believes Mexico will pay for that wall.