I failed - but it’s a good thing.
Using ChatGPT to write dialogue
For those of you who have been following The Brief you’ll know that it’s been some time since I last posted - normally I post once or twice a week, but it’s been nearly three weeks - so what gives?
In short, I had finished up the concept art for the Seraphim, the Ostinatum, and a new main character Mother (which I owe you) and then, brimming with the confidence, I went down a long rabbit hole - and face planted.
But as you’ll see, failing is a good thing.
In the early stages of a startup, a key focus is developing a minimum viable product (MVP). This concept was also central to my filmmaking process. To determine my MVP, I adopted a method akin to Leonardo da Vinci's approach to studying the human body, which heavily relied on experimentation. After considering both a trailer and a 10-minute short film, I decided on the latter for my MVP.
The choice of a short film over a trailer was deliberate. I believed a trailer, typically comprising brief cuts and montages, wouldn't provide a comprehensive test of our capabilities. My objective was to craft an impactful opening scene for "Shepard’s Tone" that met several criteria: 1) It needed to be engaging and captivating right from the start. 2) It had to effectively set up and deliver on the narrative expectations I envisioned for the opening. 3) The scene required a range of cinematographic elements, including diverse shot types, sizes, camera movements, and staging, to adequately assess the potential of AI and Unreal Engine. 4) It was important to incorporate a variety of scenes, ranging from action sequences to close-up dialogues, to test their versatility in different cinematic contexts.
So I began.
I hypothesized that by providing ChatGPT with ample context, it would be able to generate a script draft suitable for further refinement and iteration. To test this, I fed the chat detailed information about two main characters - Phi and Mother, representing the protagonist and antagonist. Additionally, I included background details about Aegis, its history, the seven districts, and the Lunaris Tower. I also described the objectives for the introduction and the desired payoff. With all this context in place, I instructed ChatGPT to write a 10-minute short film script.
For a solid week, I tried to stick to this hypothesis and test it from every angle - yet nothing worked. Every single time, ChatGPT produced absolutely terrible writing.
My hypothesis was false - I had failed.
So I sat down and thought about why my experiment had failed and I realized something quite profound.
THE CONTEXT AXIOM - When a human writer writes a scene, they not only have the purpose of the scene, the context of the story, the plot, the character and location bibles in mind, they also have their entire life experience in context.
To put this in context, a human in the natural act of writing can access and hold in context every book, every movie, and experience they’ve ever had in memory and use that context as a tool to write well. For the average human, or rather the average human screenplay writer, that would be massive - think on the order of millions of pages of data.
In contrast, at the time of my experiment, ChatGPT could only hold 2-3 pages of context. And even with the new ChatGPT Turbo which they heralded as the coming of a new age, ChatGPT can now hold a whopping 300 pages in context.
That’s not enough for my experiment to succeed.
So I recalibrated my tests and optimized for success given these constraints. The trick was finding the line of likely success. So I put together a list of possible tests and calibrated, given the constraints of 300 pages of context, where the success line would fall.
Write a screenplay
Write a 10 minute sequence (confirmed as failure)
Write an individual scene
Write a partial scene
Write a small interaction within a scene
Write three good lines of interconnected dialgue
Write two good lines
Write one good line of dialogue
To find the success line - I worked backwards.
Can ChatGPT write a single good line of dialogue? The answer - yes.
Here’s how I did it. I did all of the preloading I mentioned before, and then I also preloaded a chat with the scene I was writing and asked it to right THE NEXT LINE of dialogue. In fact I’d ask it to write 10 variations of the next line of dialogue.
Boom - it worked.
To be specific, in the opening scene of Shepard’s Tone, a Seraphim has died and Mother (the High Priestess) is giving a sermon. I wanted the dialogue of her sermon to include a call and response with the congregation (similar to many religions). So I preloaded ChatGPT with all of the character, location, and backstory bibles, and then I wrote the first part of the scene and fed that into ChatGPT as well. Then I explained to ChatGPT that it’s been 10,000 years since now and a new religion has been formed - I preloaded the chat with the beliefs of that religion, which ChatGPT helped me creat, and then I asked ChatGPT to come up with the first “Response” of the call and response.
I’ll be sharing the script soon, but I’ll say in short - WOW! It was really good.
So then I tried the same experiment and asked ChatGPT to write the next two, then next three call and responses. WOW! It was really good.
And then I tried to say, now write the rest of the sermon. AND WOW DID IT SUCK!
The Success LINE - I realized what it was - if I asked ChatGPT to only look ahead and write a little bit of the dialogue that would immediately happen next and with instruction - it did a good job - but the further I asked it to postulate the dialogue over a longer period of time and with less direction - it got dramatically worse.
This then led me to another realization, or rather hypothesis to test. What if I used ChatGPT for micro-decisions when it comes to writing dialogue - e.g. what would this person say to that person in this context sort of thing. And boom, I was moving again. Slowly, painfully, but it was now working.
After three days of doing this I had my first 10 minutes written.
That was when I had my second big discovery. Could I use ChatGPT to analyze what I had written and critique it?
No. I tried 100s of times and from every angle. Failure again.
So I thought, what if I inverted the process? In other words, I asked it to write everything but the dialogue and then use that result as a scaffold to analyze the dialogue I had, one line a time, written with ChatGPT. So, I fed ChatGPT what my set up was, my desired payoff, and the style of intro I wanted and then INSTEAD of asking it to write the scene, I asked it to write ten possible structures with demarcations for key moments, exposition turns, and pacing. I also asked it to suggest editing structure that would properly match the pacing and to suggest shot types, camera movements, shot sizes, and framing.
The result was kind of magical. While ChatGPT was terrible at analyzing dialogue, it was great at postulating pacing, cinematography suggestions, and edit points.
Then I realized something - The current workflow for creating a movie is a sequential line of of silo’d functions that a director attempts to use to describe how they want the film to be shot and edited - the script, the storyboard, the concept art, the shot list, and editing. And this is because in the current way of doing things, all of these jobs are individual people with silo’d skill sets and functions - e.g. specialists.
A THEME TO REMEMBER - In a world of AI, can a generalist, augmented by AI, combine all of the currently silo’d functions of movie making into one person and redefine what a screenplay is? - e.g. it becomes all of these things.
The net, net was that I realized that the format of a traditional screenplay is antiquated. It is only the writing - has no directorial notes, cinematography directions, or edit points. But in my world, I’m am all of the roles - so I thought…
Why not re-invent the concept of what a screenplay is from the ground up?
The final results was a ten minute screen play that is a combination of screenplay, concept art, storyboards, shot lists, and edit points. I think of it as the instructions that would ultimately be sent to another AI to then generate the film.
I can’t wait to show you. But first, I need to rewrite it 10 more times because I’m anal.
In the meantime, I’ll try to post the concept art for the Seraphim, the Ostinatum, and Mother.
What is The Brief and Who should read it?
I release a weekly digest every Friday, tailored for professionals ranging from executives to writers, directors, cinematographers, editors, and anyone actively involved in the film and television domain. This briefing offers a comprehensive yet accessible perspective on the convergence of technology and its implications for the movie and TV industry. It serves as an efficient gateway to understanding the nexus between Hollywood and Silicon Valley.
Who am I?
I'm Steve Newcomb. Functionally, I’m a recovering Silicon Valley founder that is finally old enough to have a bit of care. I’m perhaps most recognized for founding Powerset— it was the largest AI and machine learning project in the world when I founded it. It was later acquired by Microsoft and transformed into something you might recognize today - Microsoft Bing. Beyond Bing, I had the privilege of being on the pioneering team that witnessed the inaugural email sent via a mobile device. My journey also led me to SRI (Stanford Research Institute), where we laid the groundwork for contemporary speech recognition technology. Additionally, I was a co-founder of the debut company to introduce a 3D physics engine in Javascript. I've held positions on the board of directors and contributed funding to massive open source initiatives like NodeJS and even the largest such project, jQuery. My experience extends to academia, having been a senior fellow at the University of California, Berkeley's engineering and business faculties. Recently, I ventured into Layer 2 internet protocols and assisted a company named Matter Labs in securing $440 million in funding to bolster their endeavors.
What am I doing besides writing these posts?
Typically, I allocate a year between groundbreaking ventures. My exploration for the upcoming project commenced in May 2023, and the sole certainty is its nexus with the film, television, SMURF, and AI domains. Sharing insights on my research endeavors helps me discern between feasible prospects and mere illusions. My hope is that for this venture, I appropriately consider the ethical and sociological repercussions.
If you are interested in contacting me, being interviewed, being helped, or yelling at me, my email is steve.e.newcomb@gmail.com.