Professionalization is the Mind Killer

A short review of To Live and Die in L.A and more LLM writing experiments

I published a book of short stories titled Thinking About Leaving. They move between my time growing up in India and moving to the United States. It is about a hundred pages and people typically breeze through it on their travels or during a picnic in the park.

---

They don't make crime movies like they used to, partly because the nature of crime has changed. In William Friedkin's 1985 thriller To Live and Die in L.A, which I watched recently, there is a sequence that shows antagonist Rick Masters counterfeiting notes set to a synth-heavy Wang Chung soundtrack. It is a scene that kicks the movie into higher gear, setting the tempo for the action about to unfold. If you were to write the movie in 2024, you'd have to replace counterfeiting with a phishing scam or laundering gang money into betting on sports (such as in Uncut Gems). But neither phishing nor betting are acts that look good on film and require a lot of exposition with dialogue. Criminals don't work with their hands anymore. Another lost tradition.

I was surprised to learn that To Live and Die in L.A was made for a modest budget of $6 million dollars. From interviews, it looks like Friedkin employs a similar strategy to César Aira when making movies – prancing forward while making up the story on the go without a screenplay to fall back on. The effect of this is palpable in the liveliness of the movie, and in the words of one of the lead actors, "makes everything feel more real." For Friedkin, making a movie has to involve taking some risk, such as getting "consultants" to print out believable counterfeit notes, filming with non-union crew members, and making up the story while filming – all with a budget that probably won't be enough for one episode of a prestige TV show these days. To Live and Die in L.A is another example of an amateur approach to making art and taking risks paying off. Professionalization is the real mind killer.

One of the other things that stood out from the making of the movie was Friedkin's description of a car chase as pure cinema – something that can only be faithfully captured in movies and not in any other medium. Every advancement in a medium seems to come with the development of a unique vernacular. During the impromptu Summer of Protocols writing workshop (which I'll get to next), Venkat talked about how impressionism was a response to the advent of photography. The camera made it possible to capture moments in time, alternate angles, and lighting, which was then subjectively interpreted by impressionist painters. The SCP wiki, a collaborative writing project with idiosyncratic rules , consists of a set of procedures that can only exist because of the medium of the internet, which is a big reason why it produced There is No Antimemetics division – arguably the best science fiction of the last 10 years. Writing using LLMs will truly arrive when there are early signs of a unique grammar being formed.

All of the above makes me think that there should be a corollary to the popular AN Whitehead quote

Civilization advances by extending the number of important operations which we can perform without thinking of them.

Applied to art, it becomes:

A medium advances at the rate at which procedures exclusive to the medium can be generated.

When that stops happening, mediums ossify. People start carving out fiefdoms that disguise themselves as scenes advancing the art. I think the crabs in a barrel effect that you see on places like Substack, where practitioners of a medium are resistant to new technology, is a side effect of ossification of the medium of writing – particularly fiction writing, which I would argue has not changed much since the post second world war years that produced cut-ups, beats, golden age of sci-fi and metafiction.

More AI Writing Experiments

In the previous issue, I wrote about some AI writing experiments and inspiration for them.

Last week, a couple of us from Summer of Protocols, along with several participants, took part in a 5 day writing workshop. The prompt was to write protocol fiction. Since we were short on time, we thought it would also be a good place to practice writing using an LLM.

I was pretty happy with my output. I had been reading about the Oulipo – a writing movement founded by Raymond Queneau, which consisted of writers and mathematicians experimenting with writing using arbitrary constraints – such as avoiding the letter e or writing the same story in 99 different ways. So, I wanted to inject either chance or arbitrariness into my text.

Oulipo examples from ChatGPT

At the same time, I remembered a little YouTube clip of SouthPark writers talking about how to tell a good story apart from a bad one. According to them a good story would have beats that combine and, but and therefore. If a story is just event after event with no connection or payoff in between, you end up with a sequence of events and not a story. A good story would have the structure - I did something and expected something to happen but instead something else happened and therefore I ... and so on.

So I wrote some back stories about two characters I came up with and asked the LLM for suggestions, which did not yield much, but I primed it for writing the actual story. Then I wrote one passage and ended it with and, but or therefore, and asked the LLM to write the next passage.. About two passages in, I realized that if I decided which of the three adverbs went at the end of the passage, there would be very little surprise, and I would be exerting control over the story. So, I used a random number generator and assigned a number between 1 and 3 to the adverbs. At the end of every passage I used the random number of generator to decide which adverb would go at the end of the passage, and the LLM picked it up from there and wrote the next passage.

There were points in the writing process where the LLM did not move the story forward and kept adding more fluff. I had to redirect it to move the plot forward when this happened. I have actually left some of this in the story because it matches the metafictional nature of the story.

You can read a draft of the story here. I'll be doing a few more passes of editing before I publish it somewhere.

Everyone who attended used a different technique in their writing. Some focussed more on the LLM generating the outline of the story, and others focussed on sentence or passage-level input from the LLM. The other methods used were:

• Travis had the most technically complex method, where the entire story was written almost as a conversation between Claude and ChatGPT, copy-pasting text generated from one into the other and asking the LLM to continue the story

• Shreeda used the LLM for expository details and also the outline of the story. The outlining took multiple passes. Mike also used the LLM for the expository details

• Venkat used the exquisite corpse method where Venkat wrote one line and the LLM wrote the next one. Once again, the LLM did not advance the story by itself and needed some prompting to advance the story

During our discussions at the end of the week, we were surprised that it was hard to tell what parts were LLM written – partly because these are all unfinished drafts – but there was sense of finding a there there, during the writing process and while reading drafts by others. I’m currently riding the moat of low status that comes with trying out a new thing. I find the process genuinely engrossing so I’m keeping with the experiments.

The 80s is strong in this one