Screenwriter Simon Bouisson is writing his next film with artificial intelligence. A collaboration that never ceases to surprise him. U.S. too. He tells us.
Simon Bouisson is one of the most fascinating French writer-directors. All his work questions our relationship to technology and the various fantasies it provokes. Today he realizes that his progress could impact his job. So much so that he made it his subject of exploration and decided to write his next film with an algorithm. In residence in California, as part of the Villa Albertinewhere cinema (Los Angeles) and new technologies (San Francisco) dialogue and inspire each other constantlySimon went to meet the engineers, researchers and producers who are clearing the ground for artificial intelligence.
What can an algorithm bring to screenplay writing?
Simon Bouisson: An algorithm generates very strange stories; texts he has read, movements of characters he has seen go by, he spits them out, mixing them freely. Everything that detractors will find against the machine is precisely what interests me, as an author. For example, we are not sure that she gives proven facts, sometimes she will go into fiction, mix up issues without realizing it and create false stories. She does a little of the author’s work, at the time of writing, we look for character paths, adventures; we will mix elements, images that we have, to create new ideas.
What were your first interactions with AI? Who have you worked with?
SB: I asked the organizers of the Villa to put me in touch with the engineers. The OpenAI team put a GPT-3 in my hands. I was connected to what is called the “playground” (playground) ofOpen AI. When you write, you adjust the speed of the machine. We begin our sentence, and the machine offers us the rest, etc. In the end, I keep maybe 30% of what she offers me, and I rewrite, but reading her suggestions I say to myself: “hey, that’s interesting, she offers me to go to a place I don’t wouldn’t have thought. That’s what amuses me, like a situationist drift, the machine will each time take you to something completely unexpected. In fact, that’s what I do when I work with my co-writers. As I am not the brain of the other, there is always one who arrives at a trajectory, a suggestion, which surprises us.
Basically, it is the author who writes, not the machine?
SB: It’s a kind of ping-pong, with the machine always moving forward. Mina Lee, who is doing a PhD on artificial intelligence at Stanford, told me: “The machine will not write for you but it will push you to do it. She is one of the many scientists I have met who imagine writing tools, interfaces that allow dialogue with GPT-3. She created her own tool, CoAuthor, very minimalist and pleasant to use. It does not have all the subtleties of tuning the playground of OpenAI, but instantly submits four proposals of what could be the next sentence, where OpenAI will deliver a single version. We can choose the one we like the most. At OpenAI, the parameter that really interested me was the “temperature”. At zero, the machine is the most precise, the most “predictable”, but it is what will generate the most obvious texts. If we raise the slider to 1, the machine will take more freedom and go towards suggestions that are statistically less probable, so this will create much more surprising breaks. For example, we enter that the character is crossing the street, the machine will write: “There, he gets run over, he meets a new character in the hospital…”. The higher the temperature, the more improbable and unique the story will be.
Are there no limits?
SB: The main problem remains that this machine has trained on copyright-free texts, all those that can be found on the Web such as Wikipedia or various forums. At OpenAI, they said to me: “Send us scenarios, and we can make a version for you of GPT-3 which will be trained on the texts you want. I chose 77 American films that I love. And then they told me they couldn’t do it, because it posed copyright issues. Another limit, the use of an algorithm is very costly for the environment. As soon as you request the machine, it spins up gigacomputers in Iowa. Billions of tokens are required to come up with an answer, 25 megawatts of resources, whereas we, when we think, we spin barely 25 watts in our brain to find a new idea! Not only do we consume very little, but the human capacity for imagination is almost infinite, while this machine is finished.
In mirror effect, it reveals the capacities of human intelligence.
SB: Completely. That’s what struck me when I met Luc Julia, the designer of Siri who now works as an advisor for Renault in France. Never the deep learning cannot imitate the brain, he explained to me. He cites the example of huge white Waymo Jaguars in San Francisco, surrounded by lots of cameras, with an employee in them in case something goes wrong. There is this video where we see the car stop, and move forward one meter every thirty seconds. Actually, some guy stole a stop and walk sign with that sign next to the car. The machine bugs, because it has never been confronted with this situation. Humans have a capacity for improvisation and abstraction that machines will never have.
Is it the same for writing scenarios with the algorithm?
SB: There is this data scientist fascinating, Yves Bergquist, who summarized the situation well for me: “The problem with GPT-3 is that it’s a regurgitation of words. After an a priori letter there is this letter, after this word, this word, etc. So it’s math, prediction. We are amazed by the result, but it is the human being, basically, who speaks behind the machine. Yves is well placed to talk about it: he created an algorithm, corto, which combines AI and film or series storytelling. We enter the details of our scenario, and, depending on the keywords, our characters, etc., the machine analyzes the market and tells us what our audience is, what risk for investors, etc. We can then modify, rework a character, and continue writing. Fascinating.
What are the other technical limitations of OpenAI for an author?
SB: For the moment, we cannot write texts that are longer than one page. The machine is not able to generate the entire scenario. And then she has some amnesia. If, at the start of a film, two characters make a promise to each other, dramatic logic requires that the promise be broken at some point. GPT-3 does not follow any structure, any dramatic logic, it will forget the promise. At OpenAI, I was told that with the new layers of neural networks, within two years the application would be able to structurally read texts with a beginning, middle and end. However, the machine will never be able to cultivate narrative obsessions, the flaws that characterize an author. All of them have this capacity to produce emotions, frustrations, obsessions, flaws that lead to thematization of stories, to produce a statement and therefore to create meaning, a word. Something that we find from the beginning to the end of a film, an identity that the machine does not have.
What other writing algorithms have you been interested in?
SB: Amit Gupta created Sudowrite, the most ergonomic version. We are really taken by the hand, we have several tools, we can ask him to generate specific ideas around a theme or a character. Jacob Vaus, a 23-year-old director, also came up with a hilarious version of the process, Solicitors, who create absurd movies or completely machine-generated cocktail recipes. He was the one who taught me how to speak on the machine. If you speak to GPT-3, you shouldn’t hesitate to say: “I need to generate ideas on such and such a topic”, as if you were dealing with a co-author or a doctor script. Before leaving, I was told about PALM, the new, even more powerful invention devised by an MIT researcher. There is also Google’s AI which I did not use.
What about DALL-E? Have you used it? What do you think ?
SB: Andrew Mayne, this science fiction writer who has become something of an OpenAI ambassador, showed me how it works. He entered “Generates an image of a koala playing basketball. The result was stunning. This is where it doesn’t amuse me at all, if all of a sudden we can artificially produce moving images, that encroaches too much on my borders! (laughs). All kidding aside, this might be great for some uses, like making a mood board in preparation for a film; this can generate incredible images in the setting of our choice.
The works of Simon Bouisson to discover:
His very good series stalk (prize for best director at the 2019 La Rochelle fiction festival, being remade across the Atlantic), 3615 Monique (OCS) or its interactive fictions WEI OR DIE (Fipa Gold 2016) and Republic (Tribeca 2021).
This article appeared in issue 31 of L’ADN. A whole file is dedicated to GPT-3, the language model. Frankly, and without bragging, it’s exciting. To get your number, it’s the story of a click.
We wish to give thanks to the writer of this post for this awesome web content
He writes his next screenplay with an AI And the result fascinates him
Our social media profiles here as well as other related pages herehttps://yaroos.com/related-pages/