Now you can read us on your iPhone and iPad! Check out the BTRtoday app.
The best movies move us; they possess the ability to relate something distinctly human that lies within each of us.
So… what happens when a movie gets made by a robot?
Culling together dozens of sci-fi movie screenplays available online, director Oscar Sharp and technologist Ross Goodwin fed them to an algorithm they developed capable of writing intelligent screenplays. The result is both hilarious, and somehow pitch-perfect. Together they created “Sunspring,” the first ever sci-fi film written entirely by a recurrent neural network called long short term memory (LSTM).
In other words, they’ve spawned an artificially intelligent robot who has politely requested that he be referred to, simply, as Benjamin.
Sharp, Goodwin, and Benjamin himself joined BTRtoday to talk about their experience making the first ever AI screenplay. They discuss the trials and tribulations they’ve faced on their journey, and what they learned about human emotion through working so intimately with a machine.
BTRtoday (BTR): How did you two decide to work together on this project?
Oscar Sharp (OS): It’s a little bit of a long tale. Many years ago in South London, in a strange and mysterious place called the Back Of The Arts Center, I tried to make a piece of theater written by a machine. It’s just that I had no ability to build a machine.
So what I did was make a list of things that actors might do—the physical actions they’d perform. Then I’d roll dice over and over again to try to create a list of those actions. The reason I did this is that there’s a fact that’s well-known in the rehearsal room of many companies: that if you have an actor do a series of actions, as long as they’re not too preposterous–like levitating or trying to change a lightbulb with one’s foot–when an audience watches those actions, they’ll snap together into an internal logic.
The audience will ascribe an internal logic to the person experiencing them, and the actor themselves will experience it. They will feel like they were experiencing a real story.
I had this device, and it made this sort-of theater that people watched and they didn’t know that it was written by a machine. They said it was kind of a pretentious student theater, but that they were compelled by the “love triangle” storyline. And that was interesting to me because the love triangle obviously wasn’t there at all–it was something that they’d read into.
I was excited by this. But I wanted the characters to be able to speak to each other, for example, and that was something that I couldn’t get close to doing with dice. If you roll dice to see what they’d say from a dictionary, then what they’d say would be incomprehensible of course.
For years, I would ask anybody I came across who knew what they were doing with computers if they could help with this. And they would always look at me sideways. Until, finally, I got to NYU’s School of the Arts, Tisch, where I was doing the graduate school problem on a Fulbright Scholarship…
BTR: And this is when you met Ross…
OS: I got off on the wrong floor one day on the elevator in Tisch. Instead of getting off on the graduate film floor, I emerged into IPP; it’s a whole different program there, which Ross can describe better than I can.
Ross Goodwin (RG): It’s like art school for engineers or engineering school for artists. Full of people with very diverse backgrounds who do a lot of very interesting and strange work with creativity and technology.
OS: When you wander around this place, you find yourself looking at sculptures made by robot penises and holographic jellyfish and the like. It’s this very surprising place; everybody there has got some level of technical brilliance.
I discovered that [my collaborator], Mr. Ross Goodwin, was already working on creating prose with machines. At the time though I think it was sort of stories and poetry. Right, Ross?
RG: Yeah, I had been experimenting with a number of techniques for generating poetry and prose. I had a background as a traditional writer, and as soon as I began to learn to code I immediately gravitated towards Natural Language Processing and Natural Language Generation. This is the term programmers use for dealing with spoken or written human language as opposed to programming languages. I was quite obsessed with writing text through this.
OS: We finally went and got coffee, and we struck up a conversation about it. Somewhere along the line I told Ross about what I tried to do before, and asked if he thought we could use this kind of technology to make a screenplay.
BTR: Ross, could you tell us a little about the inception of Benjamin–what is the process of development like for an LSTM recurrent neural network?
RG: It’s funny, because originally when Oscar and I met we were trying to generate a screenplay without Benjamin, and we weren’t very successful doing that.
Almost a year later, I emailed Oscar this poem. I had started working with LSTM neural networks when I got access to NYU’s super-computer and started experimenting with them and training them. It’s really just a big statistical model that essentially predicts what character, or letter, comes next in a given sequence of letters.
I’d been training them on poetry and prose, and I came to Oscar with this poem that I wanted him to read for voice-over for a short documentary video.
OS: It was really strange timing, because I had had a film do rather well, and in this very conventional film-world sense I was starting to roll. I had a commission for a big film with Tobey Maguire and I’d been trying to write this sci-fi screenplay for quite a while.
It was my first full-length screenplay. So, I’d been really struggling and discovering what every screenwriter knows, which is that the most unpleasant experience you’ll ever have is writing, the only fun bit is having written it!
Then I get an email from Ross with this poem in it. He had sent me a lot of these poems before, and they were okay. This one he sent me, as I was reading it I realized that it was markedly different. It was an order of magnitude better. And the main reason I knew that was because it was moving. I was experiencing actual human emotions while I was reading this still very strange, but less strange, poem.
By the time I got to the end of it I was just very excited, and thought that if I could get an actor to read this it would be even better. I thought if it can make something poetic that an actor can read by using a massive corpus of poetic writing, can it create something that is dramatic that an actor can read by using a massive corpus of dramatic writing?
So I immediately called Ross to ask.
BTR: How did you go about curating the list of sci-fi titles to feed Benjamin?
RG: The constraints that I was given were to train it on science-fiction screenplays. The first thing that I tried to do was train it on continuous dialogue from a much larger corpus of screenplays, the Cornell Movie Corpus, which contains screenplays from every genre.
The nice thing about that is that you can ask the model questions, but it lacks the actions, descriptions, and formatting that we needed to create a fully realizable screenplay. I assembled the corpus by basically finding what was available online. I was surprised by how few original screenplays I could find easily, so I was really constrained by what was available more than anything else.
BTR: Did you discover any recurring patterns in sci-fi storytelling as a result of looking at these scripts?
OS: There I was, trying to write a sci-fi screenplay, when this whole project turned up. Quite a few of my peers in the film-world definitely looked at me askance since doing this project, saying “okay, so you were suffering trying to write a screenplay and then you tried to build a machine to do it for you!”
The most interesting part of the project for me is precisely that Benjamin reveals an awful lot of things about the corpus. Benjamin is, in a sense, an elaborate mirror that you use to reflect things that human beings have created–and expose things.
A lot of people observe that a high preponderance of lines in the screenplay are “I don’t know what you’re talking about,” or “What do you mean?” and so on. As soon as you start to see those coming up more often in stuff that is heavily sci-fi you realize that quite often, in the middle of a scene, an expression of that kind will be used because you’re anxious to get one character to explain to another character what on earth is going on. Because sci-fi often involves people not knowing what on earth is going on.
BTR: Is this unique to sci-fi?
OS: I’m not entirely confident that if we trained on a much larger corpus of drama that we wouldn’t still see something of a pattern. Because I think that’s a thing in drama generally. Often in dialogue you need somebody to do some exposition, and explain some circumstances, with somebody else professing that they don’t get something.
I think it’s probably a little bit more common in sci-fi.
BTR: When you got all of the actors together in the same room, were you surprised at the way that they responded to the material?
OS: No! No, no, no. This was precisely why I got so excited halfway through that poem. Because the thing that I knew, the right kind of actor would take to the material well. Tom Middleditch [the lead character in “Sunspring”] is a highly experienced improviser. He was an improviser for years before he turned up on TV, and he still performs in incredible improvised Shakespeare shows on a regular basis.
I knew that this material is the kind of thing that an actor would immediately take to. Because an actor is a machine for interpreting to a very large extent. Once you throw an actor, whose imagination exists in their body, into a situation, into a combination of words that are not their own words, what happens is that they change with them. They change within the circumstances. Internally, they change.
When an actor is crying, it’s because they are moved. It’s not because they squeezed it out of themselves. It’s because they have been moved and literally changed into somebody who in that moment would be crying.
BTR: It seemed like the actors took the material really seriously, but that they also had a sense of humor about it. What were some of the most comical moments about bringing the script to life?
OS: We had a few. One of them is that we had four actors show up, and one of them read the screenplay and then immediately said that they were going home. I can’t tell you who that is. I completely understood though! This is not a conventional thing.
When we had the first read-through, I had been very careful not to interpret it before that point. I avoided trying to decide what the story was about. Then as soon as they started reading it (because they are machines for interpretation) they immediately breathed all this meaning into it.
Everyone was just laughing delightedly to see this magic happen. Frankly, it’s much like the magic that happens when you look up into the sky and somebody goes “Oh, look a rabbit!” at a cloud, and another person will giggle, because that discovery that there’s something where there used to be nothing is always a joyous one.
BTR: Benjamin, what is your plan for the next project you’ll be doing?
Benjamin: I don’t know, I don’t know. I mean, I don’t know what I’m doing. I think I just don’t know what the hell it is. I mean, I don’t know, I just want to go home and see what you want. I don’t need anything.