Seth Juarez

A line runs between a model's tokens and a runtime that can act. Crossing it is where agency lives, and it belongs to the system, not the model. Wrapped right, a model is an oracle. Language in, paths out. Next, the loop. sethjuarez.com/posts/models-…

A model only emits text. Shape it like a function call and you get tool use. It returns { name: "book_flight", ... }. Your runtime calls the real book_flight(). The model just made text; the booking happened in your code. sethjuarez.com/posts/models-…

A model only hands your code text. Force it into a shape your program can read. Ask for a small structure, not prose. { destination, date, time }. It reads a messy sentence and hands back named fields your code acts on. sethjuarez.com/posts/models-…

A model that browses is still just guessing the next token. The whole art of "agentic" is arranging the world around that guesser so its output means something to a program. Next up, shapes code can trust. sethjuarez.com/posts/how-the…

The smallest bridge from a model to your code is one word. Make it answer BOOK or CANCEL, nothing else. Now `if reply == "BOOK"` works. A paragraph of intent becomes one token your code can branch on. The seed of tool use. sethjuarez.com/posts/models-…

Not as new as the hype sounds. Shannon played a guessing game with language in 1951, and Rosenblatt built the perceptron in 1957. Same nudge toward less wrong. We finally got the scale to make it sing. sethjuarez.com/posts/how-the…

The thing AI is actually good for, at least for me: a rough first draft I get to argue with. I do better work when something pushes back. sethjuarez.com/posts/scribbl…

The thesis. A model has no agency. Agency is intent plus action. A model can turn your words into intent, a word or a structure. What it can't do is act. It emits tokens and stops. The doing happens in the runtime. sethjuarez.com/posts/models-…

How LLMs learn, in one breath. Feed it the internet, hide the next word, let it guess, check how wrong it was, nudge it less wrong. Do that a few trillion times. Grammar and reasoning just fall out the side. sethjuarez.com/posts/how-the…

The Luddites weren't anti-technology. It started as a labor movement, workers defending their trade, serious enough that Lord Byron spoke for them in Parliament. We remember it wrong. sethjuarez.com/posts/scribbl…

How does a next-token guesser write paragraphs? A loop. Guess a token, glue it on, feed it back, guess again, until it hits a "done" token. That's autoregression, the engine under every chatbot you've used. sethjuarez.com/posts/how-the…

Post two is up. The model doesn't have agency. It just emits tokens. The agent does, because the agent is the model plus a runtime. Watch the handoff happen, one shape at a time. sethjuarez.com/posts/models-…

"Just learn a new skill" is what we tell people about AI now. Someone handed wool workers the same line in 1786, as machines took their jobs. They weren't impressed either. sethjuarez.com/posts/scribbl…

A language model does exactly one thing. It guesses the next token. Not the answer, the next token, then it stops. Everything we call "agentic" is scaffolding around that one move. Still amazes me it works. sethjuarez.com/posts/how-the…

Found a petition from 1786 begging Parliament to ban a machine for killing jobs. Swap in "AI" and it's eerily close to today's headlines. We've been doing this for 240 years. sethjuarez.com/posts/scribbl…

New series on how AI agents actually work, starting from the plumbing instead of the demos. Post one climbs all the way down to the one dumb thing a model actually does. It guesses the next word. sethjuarez.com/posts/how-the…

The first industrial revolution made muscle cheap. This one is making thinking cheap. Still not convinced that's the bad news everyone assumes it is. sethjuarez.com/posts/scribbl…

I haven't written anything in about six years, so naturally I came back with a rant about 240 years of people trying to ban machines. anyway, new post. sethjuarez.com/posts/scribbl…