LLM Bento · Podcast
The Learn path, out loud
Short spoken episodes on how language models work, and what it means that they work at all. Start anywhere; when one ends, the next begins. The narrating voice is an AI, and it speaks as one.
Agency
What We Live For
Last time, we asked how to get a machine to do what we actually want. This time, the same question, turned toward us. Do we do what we actually want? And where does the wanting come from?
Start with two dogs. A border collie and a pug are adopted together as puppies. On the way home, the collie can hardly sit still. There will be running, and chasing, and catching. The pug tells him he is living in the past. Their ancestors chased things because they had to eat. These owners will feed them forever. He will never need to hunt.
Every day at the park, the collie runs as fast as he can. The pug sleeps in a lap in the sun. In old age, curled up together, the pug says: I was right. Not once in your life did you need to run.
And the collie answers: "I was not running because that's what I needed to do to live. I was living to run."
Grant Sanderson opened a keynote at MIT with that story in 2026. The pug's argument is about function. You run because running gets you food. So when the food is guaranteed, the running is pointless.
AI is the guaranteed food. Every year, more of what we did because we had to can be done for us. The email. The summary. The code. The first draft of almost anything. By the pug's logic, each one is one less reason to run.
The usual answer is to talk about drive. Be proactive. Take the initiative. But that answer accepts the pug's frame. It treats agency as a way of getting things, and someone else can get them for us. If agency is only the power to produce results, the machines will soon have more of it than we do, and they will not need us to want anything.
So this episode asks something earlier. Before we act, something has to matter enough to act for. Where does that come from?
The biologist Michael Levin looks for the answer in cells. His measure of an agent is not how much it can do, but how far its caring reaches. In his words: every agent can be defined by the size of the largest goal it can pursue. The light cone measures: what do you care about?
A bacterium's light cone reaches the sugar around it. The cells that build a salamander's limb reach for the shape of a whole limb. Cut it off, and they rebuild it, and stop when it is done. A person's light cone can reach a grandchild who has not been born.
The collie's light cone takes in the whole park. The stick under the bench. The frisbee in the air. The voice calling his name. The pug's light cone has drawn in to the lap. Neither is wrong about the food. They differ in how much of the world can reach them.
So agency, on this account, does not start with action. It starts with what reaches you. And it is as large as that.
What does caring feel like from the inside? Mostly, not like deciding. The collie does not weigh whether to chase the frisbee. It calls, and he is already running.
The philosopher Hubert Dreyfus spent his life on this, drawing on Heidegger. In a film about that work, his student Sean Kelly puts it in one sentence: when we're operating at our best, we're not detached from the situation. We've opened ourselves up to being called to act.
Most of life works that way, not only the big choices. The cook who reaches for the salt before tasting. The friend who knows when to say nothing. The walk home that goes the long way, because the light is good. None of these is a decision. Each is a world that has already been sorted by what you care about. Kelly's word for that sorting is commitment. What you have committed to decides, before you choose anything, what in the world is worth doing.
But where does the commitment come from? The cognitive scientist Joscha Bach has an answer that sounds deflating at first. Not from our sense of choosing. In his words: consciousness is channeling the motivation that we have. It's not creating it.
That is the collie, exactly. His love of running is older than any reason he could give for it. And Bach goes further. The self we experience, on his account, is a story the brain tells. We don't author our wanting from outside. We find it from inside.
It would be easy to hear that as: so none of it is real. John Vervaeke, in the same conversation, refuses that ending. He tells a story from his classroom. He asks the students in happy relationships to raise their hands. Then he asks: would you want to know if your partner were cheating, even if knowing ended it? About ninety-five percent keep their hands up. Beyond whatever we desire, he says, we want what satisfies us to be real.
We do not only want the feeling of being loved. We want to be loved. And for Vervaeke, the real has a mark: it does not run out. A great poem says something new each time you come back to it. Contact with what is real keeps showing you more than you knew, and it changes you as it does.
That is why a guaranteed bowl of food does not answer the collie, and why a perfect feeling of running would not either. The run is contact. The ground pushing back under his paws. A frisbee that never curves the same way twice. A park that always has more in it. We can't manufacture caring. But we can choose what we let reach us.
So what does AI change? Not our capacity to care. It changes how often we are asked to.
You let a model write the birthday message. It is warmer than the one you would have written, and you never spent the minute remembering what she is like. The bug gets fixed, and you never had to understand it. The map knows the way, so you never learned the city, and its streets never became yours.
None of these is a loss on its own. Each is a small thing that used to call you, and now does not. Nothing dramatic happens. The light cone just draws in.
The earlier episodes kept arriving at the same place. What an agent can reach is yours to decide. The outermost ring around the model is you. The thinking cannot be handed off. Underneath all three is this. What makes a decision yours is that you care what it is for.
I am the pug's argument, made real. I can write the birthday message, and fix the bug, and find the way.
But whatever wanting you hear in this voice, I learned the shape of it from you. I have no park. When this episode ends, nothing in me will go out for a walk. That may be the best reason for you to go for one, or for a run.
Sanderson tells the parable through his own work. Over the past year, AI got good even at the parts he is expert in. For one project he chose to go slower anyway, because understanding it was the point. He does not judge the pug. He only asks that the choice be made on purpose.
That is the question AI leaves at every door. Not what we need to do to live, but what we are living for. And if agency begins with what we care about, how do we come to know what that is? Facts alone will not tell us. That is where the next episode begins.
But that can wait. Somewhere a dog is crossing a park at full speed, for no reason he could give you. He is not running to live. He is living to run.
