---
title: "It thinks before it answers"
date: "2024-09-19"
summary: "OpenAI's o1 stops answering on reflex and starts to plan — the shift from autocomplete to deliberation, and why the pause is the point"
tags: ["ai","reasoning","o1","building"]
---

Last week I asked a model a question and it paused. Not lag — a pause with intent. It worked the problem, tried a path, backed out, tried another, and only then spoke. OpenAI calls it o1. I call it the first one that doesn't answer like it's afraid of silence.

Every model before this was a reflex. You typed, it replied, fast and fluent, with the eagerness of a clever student who would rather be wrong quickly than right slowly. Useful, often dazzling, but reflexive. It pattern-matched its way to the next word and dressed the guess in confidence.

This one slows down on purpose. It plans. For a hard problem you can watch it lay out steps, check itself, throw away a dead end. The fluency drops and something else takes its place. Deliberation.

I noticed it on a gnarly bit of logic I'd half-given-up explaining. The fast models kept producing answers that looked right and weren't — plausible code that fell over on the third edge case. o1 took its time and got there. The wait felt like working with someone who reads the whole ticket before touching the keyboard.

There's a cost. It's slower, it's pricier, and for half of what I do the quick reflex is exactly right. You don't want a model deliberating over a rename. The skill now is knowing which mode the moment calls for.

But the direction is clear, and it's the one I care about. Autocomplete was a parlor trick that happened to be enormously useful. Thinking before answering is a different animal. It's the difference between a tool that finishes your sentence and a colleague who stops you and says: wait, are we even solving the right thing.

I've wanted that colleague for a long time.