Most of what we call an AI agent only comes alive when you talk to it. You ask; it answers; it goes still. Between your messages there is nobody home — no wanting, no waiting, no thinking. It is a very fast responder wearing the mask of a mind.
A creature is not like that. Left alone, it still gets hungry, still watches the door, still drifts toward the thing that caught its eye. Its mind runs whether or not anyone is looking. If we want worlds that hold characters who feel alive, this is the gap to close: a mind should keep thinking between turns.
That is a smaller, stranger problem than building a cleverer chatbot. It is not about knowing more. It is about wanting something, noticing whether you are getting it, and being honest — to yourself and to anyone watching — about why you are doing what you do.
§ 01A mind that wants something
Start with the thing a responder has and a mind lacks: something to lose. A mind runs on drives — hunger, thirst, safety, curiosity, the pull of company. Each is only a number that rises and falls (a kind of homeostasis), but together they are the engine. Nothing it does is free of them; every action is spent trying to bring one of them back down.
This is the rule of this whole site, worn on the inside. Ask of a character, as of a dragon, what does it eat? — and you are really asking what it wants badly enough to move for. A mind with no drives has no reason to do anything; a world full of such minds is furniture. Give it hunger and it has, for the first time, a stake in the next minute.
The drives change on their own, slowly, whether or not anything is happening. The mind's first job is just to notice: hunger was low, now it is climbing; safety was fine, now something has spooked it. Noticing the change is what turns a stored number into a reason to act.
§ 02Choosing what matters
Wanting is not enough; it has to find something that would help. So the world is labelled with meanings: this is food, that is water, the doorway is safety, the strange glowing thing is worth a look. Those labels are the designer's commitments — the same authored meanings the founding essay argued a world runs on — and the mind reads the world through them.
Now choosing is simple, and better for being simple. Take the loudest drive. Look for the nearest thing that would quiet it. Make that one thing the target, and move toward it — or, if the thing is dangerous, away. The mind is not weighing a hundred options in a fog; it is doing what a hungry animal does, which is going to the apple. A mind is not a list of everything it could do. It is one thing it has decided to want, right now.
§ 03Noticing when it stops working
Acting is a claim about the future: I am moving toward the apple, so the apple should be getting closer. That claim can be checked. The mind watches the distance it meant to close and asks, quietly and constantly, is this working?
While the answer is yes, the mind can stay calm and cheap — it does not need to agonise over a plan that is going fine. The telling moment is when the answer turns to no. The apple is not getting closer. Something is in the way, or the path is wrong, or the world has changed underfoot. Failed expectation is surprise, and surprise is what should wake a mind up — spend more attention exactly where reality has stopped matching the plan.
This is how a small mind rations the one thing it has least of: thought. It coasts when the world behaves and stirs when it does not. Attention goes where the surprise is, and nowhere it isn't needed.
§ 04The reason and the words
Here is the part most easily got wrong, and it is the part that matters most. A thinking agent produces two very different things, and blurring them is how you get a mind that lies to you.
The first is the reason: why it is doing what it is doing, stated in terms of things that are actually true right now — this drive is highest, that target is nearest, the distance stopped falling. The reason is not a story. It is a readout of the mind's real state, and it is either faithful or it is broken.
The second is the words: the private thought it thinks (I am hungry; the apple is close) and the public thing it might say (I am going to the apple). Words are where a language model earns its place — it phrases, it gives the mind a voice, it makes an inner life legible and warm. But it must never become the source of truth. If the model narrates "I am going to the apple" while the mind is in fact walking away from it, the words are a lie, however fluent.
The language model may phrase the thought. It may not decide the deed. The reason comes from what is true; the words only dress it.
So the rule is a hard separation: behaviour is grounded in state, and the words are grounded in the behaviour — never the other way round. This is the same no-cheating instinct that runs under everything here. A world that fakes its rules is scenery; a mind that narrates reasons it does not actually have is no better.
§ 05A mind you can open
Because the reason is a readout and not a story, it can be shown. Everything the mind is doing can be laid out plainly, moment to moment: the active drive, the chosen target, the movement, the prediction, and whether the last thought came from the model or from a plainer fallback when the model had nothing worth saying. You can watch a moving agent and read, directly, what it wants, what it is going for, why it chose that, and whether it is working.
This is the virtue the last essay prized in its notebook: a thing you can open is a thing you can trust, and correct. A mind whose reasons are visible is one a designer can debug, a player can believe, and a builder can be held to. A mind that only emits confident sentences is a black box wearing a smile.
None of this asks the mind to be clever. It asks the mind to be honest and continuous: to want something, to keep checking whether it is getting it, to think harder when surprised, and never to claim a reason it does not have. A small mind that does only that is already alive between turns — and, unlike a cleverer one, you can see exactly why.
You can watch minds that work this way running live at min-mind.com — and, because it runs on your own machine rather than a server, the mind you watch is quite literally alive on your GPU while you look at it, spending your compute to think. Ask what it eats, and for once the answer is you.