A Small Machine That Writes With Me

I wanted to make something small. Not smart, not predictive. A demo that sidesteps the standardized answers.

Just a tiny program — a little machine that writes with me, or maybe for me. Not to finish my thoughts, but to bend them.

It’s still engineered by probability. The same kind that makes up large language models, that is so efficiently good at guessing the next most likely word. But instead of chasing what’s likely, I want to tune the model toward what’s fun. Toward the strange. This machine behaved less like a feed.

As part of this approach, I’m preparing more bespoke datasets—books I love, texts that have shaped me—so the system draws from a corpus that feels personal, rather than general. The goal is to shift the tone of output toward something more intimate, more idiosyncratic.

It’s like going into a bookstore, wandering. You discover.

A lot of the recommender systems—Spotify, TikTok, Instagram—aren’t helping us discover ourselves, but rather reflect back a rehearsed image we’ve already constructed. As Rob Horning writes in I Write the Songs, platforms like TikTok don’t just guess what you like, they train you to like what they feed you. This is algorithmic conditioning, behavior sculpted into a loop. Not discovery, but feedback. Every scroll is a tiny nudge. Every loop, a tightening loop.

Maybe personality in machine learning doesn’t live in the average, but in the edge cases.

Everything online is just reflecting me back.

New Behavior / New Internet vs. Old Internet.

A weird website you didn’t ask for — but maybe needed.