Hacker News Reader

The media does a terrible job of describing AI/ML.

My roommate was under the impression that Facebook shut down their experiment wherein chatbots communicated with each other because they were scared of the consequences of the machines communicating in their own language.

For the unaware, Facebook determined it wasn’t useful to continue the exercise.

View on HN

This isn't terribly hard to achieve. E.g with a pretrained LSTM generator downloaded from some place that does neural network stuff + a homebrew text classifier (I'm fond of character-level online linear models with long n-gram ranges) that's just underpowered enough to overspecialize to a certain style, but unpredictably so. Make this poor man's model generative by choosing random combinations of characters that hit a high enough score; and then feed it to the LSTM (along with normal text samples).

The underpowered generator creates weird stuff that the good generator absorbs into its language model. Boom, you are witnessing the technological singularity. Or something.

Reply on news.ycombinator.com