An AI wrote all of David Hasselhoff’s lines in a bizarre short film

An AI wrote all of David Hasselhoff’s lines in a bizarre short film(arstechnica.com)

66 points by secretsinger 9 years ago | 28 comments

Hey, I'm one of the creators of this -- I trained the LSTMs that wrote the computer generated segments in this film. The ballet choreography was generated with a context-free grammar, because I couldn't find enough material to make a ballet choreography LSTM.

Here's our film from last year, Sunspring, which was written entirely by an LSTM trained on science fiction screenplays: https://www.youtube.com/watch?v=LY7x2Ihqjmc

It's worth noting that this year I used subtitle files rather than screenplays to train our LSTMs, so we only had dialogue rather than dialogue + action descriptions. The Ars Technica article explains everything.

krylon 9 years ago | |

Just wanted to say, that this was a very strange and very interesting experience, and now I'm excited to see what might come out of such experiments a few years down the road!

rhcom2 9 years ago | |

I'm incredibly interested in the ballet part, what was it trained on? Any more info on that part?

themadstork 9 years ago | | |

Yeah, I simply ingested words in French from a few online ballet dictionaries into a JSON object, and used a recursive context-free grammar to generate (generally) grammatically correct sentences. The result was a mixture of English and French because the dancer (Sarah Hay) requested that I add a few words of her choice in English as well.

Here's some information about context-free grammars: http://www.decontextualize.com/teaching/rwet/recursion-and-c...

strgrd 9 years ago |

Today I learned that Markov chain generation = "AI."

Shame on Ars staff for conflating this parlor trick as "artificial intelligence," and failing to explain how this rather simple process works.

For fucks sake, they find the raw screenplay of a movie like Knight Rider, regex out the crap, download a Markov chain generator off of github, and train their model on the text. And of course by "training", I mean they run a single command to process the text document, and wait. Why this is described as a "...long short-term-memory recursive machine-learning algorithm" probably has to do with the fact that Ars has their hand in promoting these short films.

gliptic 9 years ago | |

They describe it like that because it's a long short-term memory recurrent neural network, which is a machine learning algorithm.

How exactly were you able to infer that they were lying and were instead using a Markov chain from a few snippets of dialogue?

RodericDay 9 years ago | | |

defending hype-building like this is why we keep having AI winters

jlarocco 9 years ago | |

First of all, the article doesn't mention Markov, and there are a lot of ways to generate text, so you may be jumping to conclusions.

Second, the concept of AI is so fragmented, varied, and overloaded, that who's really to say Markov text generation isn't a simple AI?

And TBH, there's not much difference between Markov chains and most AI and machine learning algorithms. The math's a little different, but they're almost all based on probability and statistics (and sometimes searching) and none of them really understand anything.

themadstork 9 years ago | |

I'm one of the creators, and I assure you we did not use Markov. I trained the LSTMs myself.

orthoganol 9 years ago | |

Probably because they trained an LSTM model to learn conditional probabilities between words and then sampled using markov methods to actually generate the sequences.

It's actually pretty standard and not particularly innovative, technically, in that you can google already existing tutorials on exactly how to do that (even in Keras - forget lower level stuff).

At any rate, this doesn't work well, typically. I'm sure they generated a bunch of sample sequences and cherry picked ones that made even a little sense for inclusion in their script.

minimaxir 9 years ago | |

The tutorials of which are all extensions of the char-rnn project. http://karpathy.github.io/2015/05/21/rnn-effectiveness/

A large amount of AI-generated text just swaps in the source dataset, generates the text, calls it AI, and gets many blog posts written about "artificial creativity."

themadstork 9 years ago | | |

That's definitely how I started, but I'm working with my own LSTM now. Here's a bit about my process from last year:

https://medium.com/artists-and-machine-intelligence/adventur...

JustFinishedBSG 9 years ago | |

https://en.wikipedia.org/wiki/Long_short-term_memory

alejohausner 9 years ago |

Ok, so it's not very smart AI; emacs's eliza might do as good a job.

But setting aside the article's lack of technical rigor, and its scientism, the film itself is interesting. I was impressed by how the actors could convey feeling, even with a randomly generated script. I imagine that, in acting school, actors have to do this sort of thing as an exercise: learn how to convey feeling using random words or grunts (is this done?).

For me, it's interesting that there's another side to human communication that has nothing to do with verbal meaning. It sounds hard to do THAT using software.

Declanomous 9 years ago | |

Are they actually conveying emotion, or is the kuleshov effect [1] at work? I know one of the things they teach film actors is to 'act less' when they don't know what to do. We as humans tend to interpret the actor as displaying an appropriate emotion if their face is otherwise somewhat neutral.

[1] https://en.wikipedia.org/wiki/Kuleshov_effect

alejohausner 9 years ago | | |

Cool! I had thought that Eisenstein came up with that, but it looks like Kuleshov taught Eisenstein.

I guess that's why there's an Oscar for editing! The temporal context of a shot matters as much as the shot itself.

But even if my reaction to the film is conditioned by editing, and not just Hoff's acting, I still find it weird that I'm feeling things that don't depend much on the content of the words he's speaking.

theprop 9 years ago |

Interesting. Sidenote: a 30+ year old film Death Wish (about a woman who's dying young and how her life becomes a reality tv show) predicted AI-written novels.

krylon 9 years ago | |

Can you tell me the name of the director, producer, or some of the actors?

That sounds like a very interesting movie I would like to watch very much. But search for Death Wish only turns up tons and tons of the one with Charles Bronson going postal.

theprop 9 years ago | | |

Sorry it's Death Watch!! My fault!! Some sort of a slip! It's a bit slow as a film, but well-done and engaging. Harvey Keitel and Max von Sydow are in it. The director is Bertrand Tavernier.

beaconstudios 9 years ago |

do people actually fall for this? It's so clearly just writers passing it off as an AI because ML is the buzzword of the month that gets funding and attention.

floatrock 9 years ago | |

The overall story was written by humans. The just-barely-not-quite-nonsense that the actors speak when they get "turned on" are the lines the algorithm wrote.

If you view the credits, it shows all the corpuses the various bot-generated lines were trained on. Hasselhoff's lines were trained on a whole bunch of baywatch episodes.

Frenchgeek 9 years ago |

https://www.youtube.com/watch?v=oNyXYPhnUIs ?

krylon 9 years ago | |

Wow, I have never seen Knight Rider in English. The narrator's voice sounds a lot more dramatic than the German sync-version.

krylon 9 years ago |

Wow, this is probably the weirdest thing I have seen on the Internet so far.

But if we end up with androids dreaming of electric sheep, this is where it started.

futun 9 years ago |

Terrible.