A New Literature

Initial Impressions

Today I launched my first companion model, codenamed Rent-a-Kent. It’s a new genre of reading material, marrying a blog post & a specially trained large language model. You read, read, read, and then the model asks, “Do you have any questions or comments?”

The model started with generic ChatGPT but to that my team added substantial quantities of my own writing (I have ~5m words out there in various forms). They added the text & pictures of the blog post itself. Finally, they also added prompts attempting to keep conversation centered around the topic of the post.

It struck me during preparing this piece that we were engaged in a new form of literature. Writing the blog post itself was conventional:

• Writing words in a row

• Drawing pictures

• Designing (beneficially relating elements)

“Writing” the companion model was more like programming. I wasn’t putting words in a row that people would read in a row. I was giving instructions now that would be interpreted in context later. That’s programming.

But not exactly programming, because I felt very little control, & only a modicum of influence, over what words would come out of the companion model’s “mouth”. This isn’t the red & green of TDD. What kind of questions will folks ask? If they make comments, what kind of response do they expect? When questions are answered, do the answers make sense? Are they roughly correct or would I contradict the answers if I had the chance? Would I phrase the answers differently, perhaps by introducing a pithy analogy or by telling a story?

Impressions

I’ll know more when we’ve had a chance to review the logs of actual interaction with the post. However, I can already share impressions:

• In one way this is “just” the evolution of the comment section.

• Except you don’t have to wait for someone to reply.

• Except questions/comments in this form preclude the possibility of human connection. (Something to think about for later.)

• But also there’s no incentive to be an incendiary asshole in this format, unlike regular comment sections.

• Seems like the “training” part of the writing has a long way to go. The test questions I posed didn’t give very good answers (where “good” is “kind of like what Kent would have written”). We’ll have to see how long folks’ sessions were. We’ll be winning when people engage for tens of minutes after reading 5 minutes of prose.

• It’s a little soothing to me that I don’t feel like I have to stay on top of all the comments posted. It’d be a lot more soothing if I was more confident in the generated answers.