What the Co-founder of OpenAI Does to Automate Work

Most professionals who care about improving their work are running somewhere between 20 and 50 experiments per year. A new email subject line. A revised landing page headline. A tighter system prompt that produces more consistent AI output. Maybe a new CTA on a key form. You test one thing, wait for results, decide what to keep, and move on — because that's all the bandwidth you have.

Andrej Karpathy, the former Tesla AI director and OpenAI co-founder who coined the term "vibe coding," looked at that same loop in March 2026 and asked a simple question: what if an AI agent ran the experiments instead of you?

The answer became autoresearch — a 630-line open source script that racked up 42,000 GitHub stars in its first week. Fortune called it "The Karpathy Loop." Shopify CEO Tobi Lütke pointed it at Liquid — the template engine powering every Shopify store, a codebase he originally wrote 20 years ago — and got 53% faster rendering from 93 automated commits. Everyone covered the ML side. Most business professionals closed the tab. That was a mistake.

This is my latest project, while we do have audio summaries for each newsletter. They are not ideal for listening, they are simple text to speech. So what we did was create a way to provide a weekly summary of the newsletters in this podcast. And actually it’s a work in progress. Right now you get a pretty good podcast recap of the previous week’s newsletters. But over time they will be better. That’s the plan.

Autoresearch's three-part pattern works on anything you can score — prompts, emails, landing pages, and beyond

Karpathy's core insight is that almost all knowledge work improvement follows the same loop: modify something, measure it, keep or discard, repeat. Humans do this manually and slowly. An AI agent can do it continuously and at scale. The pattern has three parts, and none of them require a GPU or a machine learning background.

Normally, you get the option to listen to the newsletter when you click the button on the top but the newsletter is meant to be read. So I am currently experimenting with generating a weekly podcast that keeps you up to date every Monday. I’ve already published four episodes but the first couple weren’t that good. My intention is to offer this up every Monday as a way to stay on top of the AI news for the last week.

Every autoresearch loop runs on the same three primitives:

Editable asset. One file the agent is permitted to modify. In Karpathy's original implementation, it is train.py. In a business context, it could be a system prompt, an email template, a landing page, or a Claude skill instruction file.

Scalar metric. One number that determines whether the change was an improvement — computed automatically, without human judgment. Validation loss for ML training. Open rate for email. Conversion rate for landing pages. Scoring rubric pass rate for AI-generated content.

Time-boxed cycle. Each experiment runs for the same fixed duration. Karpathy uses five minutes per training run, which yields roughly 12 experiments per hour and 80 to 100 overnight.

All three present and the loop runs. Any one missing and it doesn't.

This is the highest-leverage application for most AI Enterprise readers, and it requires no GPU at all.

Four Steps: The four-step process prompt and AI Skill Optimization.

Step 1: Identify the system prompt or skill file you want to improve. Choose one that produces inconsistent outputs or requires manual editing more than 30% of the time.

Step 2: Define 3 to 6 binary evaluation criteria. Each must be answerable with a yes or no. Examples: "Does the output include a concrete next step?" "Is the response under 200 words?" "Does it avoid the word 'delve'?" Binary criteria are critical — sliding scales give the agent room to game the checklist.

Step 3: Write a program.md file that tells the agent what to optimize, what constraints to respect, and when to stop. Karpathy's own example is 40 lines.

Step 4: Point a coding agent (Claude Code, Cursor, or Windsurf all work) at the repo and walk away.

Time: 20 minutes of setup, then hands-off overnight.

Example results: Balasubramanyam Kosuri, a technical writer, published the actual run log on Medium. He applied the loop to a documentation SEO skill and went from 24/40 to 40/40 in 14 autonomous cycles — three consecutive perfect scores triggered the stop condition.

Your editable asset: A single email template file.

Your metric: Open rate, click-through rate, or positive reply rate — pulled from your ESP automatically.

Marketing teams currently run 20 to 30 experiments per year on email. Teams running autoresearch-style loops could run 36,500 experiments per year.

Four Steps: The four-step process to Email Subject Line and Copy Testing.

Time: 30 minutes to configure; experiments run overnight.

Your editable asset: The landing page HTML or a component file in your CMS.

Your metric: Conversion rate, tracked via your analytics platform.

Four Steps: The four-step process to landing page optimization.

No channel selection. Autoresearch optimizes within a channel. It cannot tell you whether to be running email campaigns at all.

No brand judgment. The agent optimizes for whatever metric you give it. Constrain tone and voice explicitly in program.md.

Requires a measurable metric. If your goal is "make this feel more on-brand," autoresearch cannot help you. You must translate your goal into a number.

The ML version requires a GPU — but you have options.

The original autoresearch repo requires a single NVIDIA GPU. Here are your paths:

Run it on a Mac. Two community forks bring autoresearch to Apple Silicon:

A Mac Mini M4 Pro (24GB+) handles either fork well. The base M4 (16GB) requires lowering the eval batch size.

Rent a GPU in the cloud. An overnight run costs $16–20:

Skip the GPU entirely. For prompt optimization, email testing, and skill improvement — no GPU needed. Just a coding agent (Claude Code, Cursor, or Windsurf) and the file you want to improve.

The shift Karpathy is describing is not about GPUs or training loss curves. It is about moving from experimenter to experimental designer. You define what "better" means. The agent runs the rounds you would never have time for.

I appreciate your support.