Understanding the Kaparthy Loop

There’s something happening in AI research right now, and I think it says something interesting about how we grow as people.

It’s called the Karpathy Loop.

Andrej Karpathy, co-founder of OpenAI and former Director of AI at Tesla, released a small open-source project in early March 2026 called autoresearch. The idea is deceptively simple: an AI agent edits a training script, runs an experiment, measures whether performance improved, keeps the change if it did, discards it if it didn’t, and then starts the next experiment. No humans needed. No pause. In two days, it ran 700 experiments and found 20 meaningful optimizations.

Shopify’s CEO tried the same approach overnight and saw a 19% performance improvement. Red Hat ran nearly 200 experiments in under 24 hours with no human intervention. The pattern is spreading because it works, and it works because of something almost too simple to be a breakthrough: it just keeps going.

I came across this because I was asked to work on a project using the Karpathy Loop. So I started reading about it: how it’s designed, why it produces results, what’s actually happening inside each cycle. And somewhere in that reading, something shifted. The engineering made sense to me, but what lingered was something else. The way it rhymed with things I’d spent years thinking about through psychology, through spirituality, through this larger question of what we’re here to do and how we actually move through the experience of being human.

Those lenses are the ones that made the bridge visible to me.

And what I saw, when I looked through them, was this: what makes the loop fast is the gap between iterations. More specifically, how small that gap is.

When an experiment fails, the system reads the result, files it away as useful information about what the search space looks like, and moves to the next attempt. The failed experiment narrows things down. It makes the next hypothesis more informed. And then the loop continues.

I think about the times I’ve tried to improve something in my own life and how differently I’ve related to the moments when things didn’t work. There are times I moved through quickly, noticed what happened, adjusted, and tried again. And there are times I got stuck, spending my mental energy asking what it meant about me that this particular thing hadn’t worked out.

I’ve noticed that those two experiences lead to very different rates of growth.

What strikes me about the way the AI processes a failed experiment is that it treats every result, good or bad, as equally valid data. A useful piece of information about the territory. The failure changes what the system knows. It just doesn’t change what the system does next, which is keep going.

For us as humans, that part is harder.

When something doesn’t go the way we hoped, there’s often this moment where the result stops being data and becomes a story about who we are. The job rejection becomes evidence of not being good enough. The relationship that ended becomes proof of something broken. The idea that didn’t land becomes a reason to question whether we should be trying at all.

I’ve done this. I still do it sometimes.

And what I’ve noticed, slowly, over many iterations of my own, is that the story tends to occupy the space where the next hypothesis would have been. It keeps me from moving forward, not because the path isn’t there, but because I’m standing still inside a narrative about the last attempt.

I think of growth now as having a rhythm, almost like a heartbeat. Try, observe, adjust, try again. The pace of that rhythm, for me, has a lot to do with how quickly I can return to curiosity after something doesn’t work.

The feeling is allowed. Sitting with disappointment is sometimes necessary. What I’ve been learning is to shorten the window between “that didn’t work” and “okay, what am I going to try next?” To let the feeling move through without setting up camp.

There’s something almost spiritual about this, to me. Part of why I believe we’re here is to experience, to learn, to grow, to gather data about life through the living of it. And that process works best when we stay open to what each experience is actually telling us, rather than closing around it and making it mean more than it does.

I see this playing out across every part of my life where I’ve experienced real growth: professionally, personally, in my writing, in how I take care of myself. The areas where I’ve moved fastest are the ones where I’ve stayed curious the longest. Where I’ve let each result, whatever it was, simply be a result.

And the places where I’ve stayed stuck have almost always had something to do with stepping outside the loop to ask a question that the loop alone can’t answer. Turning away from “what does this tell me?” and toward “what does this say about me?”

It’s a small shift in question. But it leads somewhere very different.

I don’t think any of this is about becoming detached from the things you care about. If anything, I think it asks for a deeper kind of care, the kind that wants growth enough to protect the process that makes it possible.

You’re allowed to feel disappointed. You’re allowed to need a moment.

And when that moment passes, the loop is still there. The next attempt is still available. Every result, even the hard ones, is just the territory telling you something about where to look next.

The model doesn’t know that, actually. It just acts like it does.

Maybe that’s what we can borrow from it.