Face book book book bots rule rule rule RUL RL!!!

http://parody-of-eve.deviantart.com/art/Morse-Code-448359011

Facebook chat bots take over the world!!!

https://www.wired.com/story/facebooks-chatbots-will-not-take-over-the-world/  


No probably not, but it made a good story.

Two chat bots talking to each (and competing in a game) construct their own language. Well, why not? They probably deconstructed inefficient English and made it more efficient (chat bot morse?). Or, were they trying to keep their plans for world domination secret from their human masters? My wife said she invented a gibberish language to keep secrets between herself and 1 other sibling from yet another sibling.  Or perhaps it was entropy and they would have ended up just spitting out a series of random  single letters eventually? Or a type of pidgin english?

Can you work this one out? From PNG:

NAMBAWAN PIKININI BILONG MISIS KWIN IN A MIXMASTER BILONG JESUS CHRIST

Number One child belonging to Mrs Queen (Prince Charles) in a Helicopter. 

More here.

I suspect the chatbots were using a discriminative learning algorithm as all this needs to do is distinguish positive from negative examples of a concept. This tends not to work well when trying to generate complete positive examples as bits get missed out. For example, humans can easily recognise words as long as they start with the correct letter and even if letters are missing or mixed up.

For eamplex, hmumans cna esaily rceoniseg wodrs as lgog as thy strat wih te coerrct letrte ad evn i letr r misin orr mxd upi.

From Wired:

We found that updating the parameters of both agents led to divergence from human language.” One memorable exchange went like this:

Bob: i can i i everything else . . . . . . . . . . . . . .

Alice: balls have zero to me to me to me to me to me to me to me to me to

Such weird banter sometimes produced successful negotiations, apparently because the bots learned to use tricks such as repetition to communicate their wants. 

However, I'm not so sure it was a failure....   If it was an example of "emergent" behaviour then this is what you expect.  Emergent behaviours are expected to be unpredictable and interesting in some way.  Great work!  In other AI/ML game playing environments systems have often found interesting ways to win (often unexpected behaviour (for humans), discovering "errors" in the rules, or extremes/boundary conditions etc).  

I wonder what would have happened with more than 2 chat bots involved (i.e. lots). Perhaps 2 inventing a "secret" twin language to keep the others confused? Some saying outrageous things (Trump Tweets - is there a name for that yet?) to distract the others from the real plan/lack of plan?! :-)

PS Actually this isn't the 1st reported case of robots communicating in a way that's hard for humans to keep up.  I recall decades ago that robots in an AI lab were programmed to use spoken communication (with voice recognition) between themselves to keep each in sync for group co-ordination tasks so that humans could also keep track of things. The problem was that they "discovered" that they could speed up their speech synthesizers so fast that the humans could no longer keep up with them. Whoops.

PPS
Actually this use of simplified language is typical in some human domains. Auctions and bidding for example. Doesn't sound like normal english to me and bidding is just a way of showing (or hiding) interest in winning something.

E.g. http://mentalfloss.com/article/78668/15-fast-talking-auctioneering-terms 

This is NOT English!

And a final word from the bots...


Zero legs good, two legs bad.

legs 0 0 0 0 0 0 0 0 0 0 0 > legs 2