The Karpathy Loop, Applied to Marketing

Google will optimize your ads for you. Inside their platform, using their algorithm, measured by their definitions. You don't see how it works and you don't get to change what "better" means. You get a dashboard that says trust us.

Meta does the same thing with a different name. Advantage+ picks the creative. Performance Max picks the audience. Both call it automation.

I used to call it that too, until I noticed what it actually resembles: a casino automating your betting strategy. The house built the algorithm. The house defines what winning means. The house optimizes for the house. You're not the customer in that arrangement. You're the supply.

In March 2026, Andrej Karpathy released a repo called autoresearch. 630 lines of code, one GPU, one file the agent can edit, one file only the human writes, one metric that decides what lives and what dies. The agent forms a hypothesis, edits the training code, runs a five-minute experiment, checks the score, keeps the change or rolls it back, then loops again. 126 experiments overnight, no human in the loop for any single one of them.

28,000 developers starred it in a week. Most of them saw a machine learning tool.

I saw the opposite of a black box, which was an uncomfortable thing to notice, because I'd been living inside one for years without questioning it. Just reading the dashboards it handed me.

Here's what actually makes autoresearch different from platform automation, once I sat with it long enough to name it precisely. The human defines the constraints. The human defines success. The human sees every experiment, every result, every decision the system made and why. The agent does the work. The human owns the logic. Nothing in it is hidden, and nothing is quietly optimized for someone else's revenue. The whole system fits inside your own head, which is not how your ad platform works, not how your attribution tool works, not how a single vendor in your stack works.

Every tool you rent is someone else's opinion about your business, repackaged as a fact. The dashboard says you're winning. The P&L says something else. You can't explain the gap, because the logic that produced it lives in someone else's box.

So I built a tool called LoopKit, and I mapped it onto Karpathy's architecture almost one to one. In autoresearch, there's a frozen file called prepare.py the agent can't touch, defining the data, the rules, the environment. In LoopKit, that's the brand brief: product, audience, voice, the constraints the agent works inside, and they're mine, not Google's. In autoresearch, there's an editable file called train.py the agent rewrites every iteration. In LoopKit, that's the ad copy itself, rewritten every cycle, angles and hooks and structure all fair game. In autoresearch, one metric called val_bpb decides everything. In LoopKit, that's a composite score I define myself: relevance, clarity, urgency, brand fit, scored transparently instead of hidden behind a number that means whatever the vendor needs it to mean that quarter.

Same architecture, different domain. The one difference that actually matters is who owns the loop.

"AI writes ad copy" is a commodity now. Every tool does that. What almost nobody does is run fifty experiments on that copy, let an agent hypothesize and fail and learn against a fixed standard across dozens of cycles. Asking AI for a draft and running an autonomous research program against your own metric aren't the same kind of act, even though they look similar from outside. One is a question. The other is a study.

I think the actual product here isn't the pattern, even though the pattern is what's transferable to email subject lines and landing headlines and bid strategies, anything with a bounded search space and a scorable outcome. The actual shift is smaller and harder to see: it's the move from renting intelligence to building it, one verified experiment at a time.

Two organizations can look at the same market. One logs into a platform, reads a dashboard built by someone else, makes decisions filtered through that someone else's priorities, and scrambles every time the algorithm changes underneath them. I know that morning routine. Open the laptop, check the dashboard, scan the green numbers, close the laptop, learn nothing.

The other runs its own loops, defines its own metric, owns every experiment, and already has the data to measure an algorithm change independently the day it happens.

The first is a customer. The second is a competitor.

I don't think the interesting question left is what AI can do. I think it's what loop you should be running overnight, on your own terms, with your own data, measured by a definition you actually wrote yourself. Not inside their black box. Inside yours.