The clickbait, hype and fear around AI (and facebook, a bit)

A short time ago, Facebook developed bots that were able to bargain, they started talking weird, Facebook developers probably giggled and re-wrote the code.

Then, the Internet exploded. TechTimes write about “stopping bots before they evolved to Skypet”. NYPost called it “creepy”. CNet mysteriously wrote that “Facebook declined to comment”. BGR wrote a hart-breaking story about panicking engineers.

To be short, internet started buzzing (and it is a shame that big news agencies joined the buzz) about something like that:

When what really happened was that:

According to real Facebook code post and pretty good explanation of Daily Mail Facebook studied negotiation process using natural language (by the way, the very source of this “skynet” is available on github). The ultimate goal was to create a bot that could negotiate — so he like has 3 apples, you have a box of matches, you say, hey bot, wanna have matches? give me three apples, and the bot answers no, max one apple, you say nonono, not enough apples, can you give me two apples, and I give to you the matches, and the bot be like heeey, it’s a deal!

Then, the Facebook team utilised a pretty simple approach to enhance the bots — they have built two bots and make them communicate. Google’s famous AlphaGo was taught the same way — it played against itself, a lot. After some time the researchers returned and saw that the conversation above was reduced to something like:

Bot A: you apple apple apple me matches give apple apple apple give matches

Bot B: no give matches give apple

Bot A: no give apple apple give matches

Bot B: ok thx

So bots used a relatively simple subset of English language to communicate, and that looked a bit weird. The actual issue that the bots were optimised to make bargaining effective, and not to keep it clear for the human. Was it a source of Skynet? No, it was a design error where “good English” was no optimisation criteria for AI system.

The problem was that while current AI/ML industry being relatively new, there are already a lot of neo-luddites who dwell on fear, mention Elon Musk with his AI predictions (which he did not make the way he is being quoted), spread the panic and make people reject the technology before it has been developed and matured(and, of course, call everything the don’t understand “a Skynet”).

This story is a open letter to all Data Scientists and Machine Learning specialists: make them stop.

Spread the word about data science. Explain to people what it is and that you are not building Skynet. Teach and coach people about the technologies, and, most important — try to make the technologies accessible to everyone.

And don’t share the clickbait. Ever.

I write about things I wonder about