michael-dean-k/

On Monday 6/15, I'm hosting a workshop to kick off a reading group for classic essays: RSVP here.

← all posts

Writing is recursion

· 411 words

I wonder if AI sucks at writing, not because our LLMs aren’t big enough or trained enough, but because engineers don’t actually understand what writing is. Even someone who has spent 1,000s of hours writing prose will still have difficulty telling you how it’s done. It’s not a logical process you can understand or recreate. Paragraphs come together, accidentally, often without method. Every person is different, and every essay is different; maybe that’s because the Idea is what’s really in control, and the writer is merely the vessel, deploying all its varied skills, unconsciously, to take form.

While I’m very confident in how to analyze finished writing, I have no system for producing it. But as I was making coffee this morning, something clicked. While there is definitely a high degree of randomness and unpredictability, there’s still a “mode” of writing that can be emulated, and it’s very different from how our chatbots work.

Human writers pause, back track, delete, and redefine their goal. Every sentence is a reflection point, where they consciously and unconsciously make decisions. Of the 100s of possible lenses, they might be consciously beholding 3, and have another 15 unconsciously operating.

A great writer is nothing but a master of recursion. They understand, at any given instance, which decision lenses need to turn on and off. Even a writer with no plan, no goal, can produce something great in a session if they’re good at recursion.

AI doesn’t do this. It (1) thinks really hard for one second to understand the assignment (by mapping what you’re asking it onto everything it’s ever seen, using vector calculus), and then (2) barrels forward at an unrelenting pace, making decisions that are dimensionally narrow—one letter at a time.

The innovations of o1, o2, o3, o4 are weird solutions, they’re saying: instead of letting it think for 1 second, let’s let it think for 10 or 100 seconds or more. Regardless of how much it can cram, it still moves forward like a handless idiot with no ability for recursion.

What if we could build a sophisticated writer with GPT-4, simply by building a writing app that re-prompts itself after every sentence? It would write one sentence, re-read from the top and scan all 81 patterns, re-configure its goals, and then decide to add, edit, delete, or start over. It would be expensive and take as long as a human (1 hour, not 1 minute), but it might be great.