LLMreading is TVfication of text

I now often prefer LLM glosses of many texts even when I don’t need compression or ELI5ication. It’s because LLMs preserve the substance while presenting it in a familiar voice I’m already used to.

This is a bit like how TV is easier to consume in larger quantities than movies because you only have to get invested in a cast of characters once to enjoy their stories for hundreds of hours. And other things being equal, I vastly prefer TV and extended universe franchises to stand-alone movies or textual fiction. In text, I vastly prefer expansive universes like Discworld and trilogies to standalone novels. In nonfiction, I’m more willing to read an author I’ve read before, like Francis Fukuyama, though the effect is less strong in scholarly writing since voices are more homogeneous. Even in literary fiction I prefer the rare cases where characters repeat, like Updike’s Rabbit series.

This is I suspect an other-modeling bias. I have low-other-modeling bandwidth and patience. I’m fundamentally not that interested in people as unique individuals, not even myself, which is why I traffic so much in coarse archetypes. I can individuate when it matters practically, example managing someone, but fundamentally I don’t think there is much that’s special about anybody. We’re all a lot more interchangeable than we think. Which means most differences feel irrelevant and mechanisms that discard differences act as noise filters for me. Reusable fictional characters and nonfiction authorial voices are a huge reduction in cognitive burden for me.

Of course this has a downside. It’s not that much fun to watch Matt Damon play Matt Damon everywhere from Ancient Greece to interstellar futures. But within limits if the specificity of the particular human is not that salient, I’ll take the most familiar substitute that does the job. This is why even with AIs, I’ve mostly stick with ChatGPT for writing. Even with model updates, it’s the same character. Claude is a noticeably different character I don’t know as well, and prefer to use mainly for code.

This is a kind of intersubjective laziness of course, and I do try to “challenge myself” by deliberately trying to encounter new types of humans and even AIs. But I think being lazy this way is overall a good thing.

I think this suggests a prediction. People who prefer TV, movie franchises, and series fiction will like reading both original generated content and register-shifted derived content more than people who enjoy one-off movies or books. The literary vs mass dimension is not that relevant. What matters is the amount of content investment in an authorial voice or character gets amortized over. But people who deeply value (overvalue?) the uniqueness of each human for its own sake are going to fare worst in dealing with AI.