Notes - Dramatron
I experimented with Dramatron from DeepMind.
- Hyerarchical approach of starting from a logline, to outline, to writing individual scenes
- Can be run from a Notebook (once you have Vertex AI running - the whole process made me appreciate the focused approach OpenAI has for ChatGPT)
- Potential improvements
- Do not rely on summaries of previous scenes - While summarization is a strong suit of LLMs, I believe that summarization is not a core part of the writing process. Ultimately, a screenplay is about characters pursuing their goals in the immediate scene. They aren’t summarizing their life story in their head; they are acting and reacting. As a writer, we may use summarization to pick the scenes to include, but that is more equivalent to picking from hundreds if not thousands of possible scenes. Feeding a scene summary to an LLM is asking it to pick the next probabilistically, it is taking one of the most powerful tools of a screenwriter and reducing it to the style found in a children’s book. If an LLM is going to write a great screenplay, it is going to have to do more than summarize and continue (which probably points to issues with an LLM outlining a screenplay as well)
- Do not rely on a theme - Similar to relying on summaries but I can lean on David Mamet in his Master Class on dramatic writing holds up a script, shakes it, and says no theme falls out of it. The only thing in a screenplay is the characters taking action in pursuit of their goals. I’m not saying that there isn’t a theme, but we need to be careful not to conflate the appreciation and study of art for how art is created.