Next Time Around

Lately I’ve been trying to think in loops. Most of it comes from work, where a growing share of what I do runs as recurring AI automations, jobs that fire on a schedule, do their thing, and come back around next week to do it again. The AI world is full of loops right now. Reinforcement learning is a loop. Every lab, and probably a few hundred startups, is trying to build the self-improving one, the system that runs, checks its own output, and comes back better.

But the idea is a lot older than AI, and the longer I think about it the more places I find it. I’ve written before about living by the seasons, which is a loop with a one-year period, and about fractal thinking with AI, which is a loop that keeps recursing into itself. Both of those posts were circling something I didn’t have a name for yet. I think the name is just loop, plus one detail that decides whether the loop helps you or traps you.

The word has a bad reputation

For most people, me included, “loop” is a negative word. You get stuck in a loop. A conversation goes in circles. The same argument with the same family member plays out at the same holiday every year, line for line, and everyone leaves feeling the way they felt last time.

The canonical image is Groundhog Day, Bill Murray waking up to the same February 2nd forever. That’s the loop as a trap. But there’s a less famous movie that gets it better. In Edge of Tomorrow, Tom Cruise plays a soldier who dies on a beach fighting an alien invasion and wakes up the day before, again and again. At first he dies within seconds. Then he starts paying attention: where the first enemy comes from, which step gets him killed, what to do differently next time. Each death is a review. Each morning is a new run with the adjustments. Eventually he’s the best soldier on the beach. Nothing about the day changed. He did.

The step at the end

So the loop isn’t the problem. The question is whether there’s a step at the end where you look back at the pass you just ran and decide what to change before the next one.

The gym is the easiest example. If you do the same workout every session, same exercises, same weight, same reps, you’ll improve for a while and then plateau. Your body adapts to the stimulus and stops needing to change. But if at the end of each session you ask what am I adjusting next time, a bit more weight, a different rep scheme, a new exercise, then each session becomes the input to the next one, and the routine improves itself. This is progressive overload, and it’s old. Milo of Croton, the ancient Greek wrestler, supposedly built his strength by carrying a newborn calf every day until it was a full-grown bull. The calf did the adjusting for him.

Cooking works the same way. I’ll make a recipe, taste it, and before I’ve finished eating I’ve already made the notes: more salt, less garlic, pull it off the heat a minute earlier. Next time I run it with the changes. After a few passes the recipe isn’t the one I started with. It’s mine.

None of this is new or clever. People do it without thinking about it. What I’ve found useful is naming it, because once you see the structure (run, review, adjust, run again) you start noticing where the review step is missing.

Ruts and navel-gazing

A loop goes bad in one of two ways.

The first is the rut: the loop with no review step at all. The same workout, the same recipe, the same holiday argument. Nothing is learned from one pass to the next, so nothing changes, and the repetition starts to feel like a sentence. This is the loop people mean when they say “stuck in a loop.” Usually it isn’t the repetition that hurts. It’s the absence of any adjustment.

The second failure is the opposite, and it’s more common among people who read blog posts about mental models: all review and no run. You can turn introspection into a permanent state, always analyzing, always planning the next adjustment, never actually going back in and running the thing. Reflection has a place in the cycle, and it’s a phase, not the whole cycle. Cruise doesn’t spend his resets sitting in the barracks studying the beach. He goes back out and dies on it.

This is something the seasons taught me before I had the vocabulary for it. In the seasonal loop, fall is the review phase: harvest, take stock, look at what worked since spring. Then winter is rest, and spring is a fresh run with the adjustments. The review gets its own season. It doesn’t get all four.

Gaining from the hits

Nassim Taleb’s Antifragile is one of the books I come back to most. His distinction is between things that break under stress (fragile), things that withstand it (robust), and things that actually get stronger from it (antifragile). Muscles are antifragile. So are immune systems, and some economies, and, I’ve long thought, good software. Every bug report, every attack, every edge case that blows something up is a stressor, and a well-run codebase comes out of each one harder than it went in. The software you trust most is usually the software that’s been hit the most.

But stress alone doesn’t make anything stronger. Plenty of codebases get hit constantly and just rot. Plenty of people lift sloppily for years and only get injured. What separates the two is the loop. Something has to notice the stress, figure out what it means, and feed that back into the next version. The review step is how disorder gets converted into improvement. Without it, the hits just wear you down.

I think that’s the most practical way to read Taleb. Antifragility isn’t a property something is born with. In most of the cases that matter, it’s a property of a loop someone is running.

Loops at work

This is where AI comes back in, and where I’ve had the most success lately.

The hard part of running an improvement loop has always been the review step. It’s where you have to look honestly at what happened, which requires knowing what happened, which requires data most of life doesn’t keep. My work, conveniently, keeps almost all of it, because nearly everything I do now passes through tools that leave a record. So once a week I take my AI harness through the same question I ask at the end of a workout: what am I changing for next time? The review that used to be the expensive part is now the cheap part. The data is there, and the AI is good at reading it and proposing the adjustment.

The result is a harness that gets a little more efficient every week. It’s genuinely satisfying to watch, a system improving on a cadence, mostly on its own, with me deciding which adjustments to keep.

The same tools lower the cost of review everywhere else too. Workout logs, food notes, a journal: anywhere you’re already leaving a trail, you can now hand it to something that will read it back and ask what you’d change. The loop was always available. The review used to be tedious, and now it mostly isn’t.

Where this lands

I know this is a little self-helpy. “Reflect on your choices and improve” is not a new idea, and people are right to roll their eyes at it. But I think the loop framing adds something the self-help version doesn’t, which is the cadence. You don’t review all the time, and you don’t review never. You run the thing, and at a set point (end of the session, end of the week, end of the season) you look back, pick what to change, and run it again.

Most of life already has this shape. The day repeats, the week repeats, the year repeats. You can’t get off any of those loops. What you can do is put a review step at the end of each one.

The soldier on the beach doesn’t get out of the loop by escaping it. He gets out by using it. I’d like to think we don’t need an alien time loop to figure that out. We’re already in a few of them.