Andrej Karpathy released a project today called autoresearch. It's an autonomous system that runs hundreds of experiments, measures results, keeps what works, and throws away what doesn't. No human in the loop for each iteration. Just a clear goal, a metric, and a machine that grinds toward better answers while you sleep.
It was built for machine learning research. But within hours, people were already asking the obvious question: where else does this pattern work?
Here's the thing. That question keeps coming up, every single week, about a different tool. And that's the real story.
The conversation has shifted from "AI can't do that" to "let me ask and see what's possible this week." That's not a small change. That's a complete reversal.
The Gap Is Collapsing
Six months ago, iterative A/B testing of email sequences at scale was something you needed a marketing team for. A data analyst. A platform with five-figure annual contracts. A/B testing your follow-up emails meant setting up experiments manually, waiting weeks for results, then guessing at what to try next.
Today, the pattern Karpathy demonstrated for ML research is the same pattern showing up in tools that test subject lines, landing page copy, and ad creatives automatically. Set the goal. Define what "better" means. Let the system iterate. Keep what wins.
The tools aren't theoretical. They're here. And they're getting cheaper and simpler every week.
If you've been building micro-automations for specific tasks, you've already seen this pattern at a smaller scale. What's changing is the speed at which each new capability becomes accessible to businesses that don't have a tech department.
What This Means for Real Businesses
I was talking through the implications of Karpathy's project with Perplexity today, specifically about our scorecard marketing system. We use scorecards as a way for prospects to self-assess before a conversation. They answer questions across several key areas, get a results page showing where they're strong and where they're exposed, and then we follow up based on what they told us.
That follow-up is where the real opportunity lives. Only 2-5% of people who engage with your content will contact you on their own. A proper follow-up system pushes that to 10-15%. But until now, optimizing that follow-up meant manually testing different subject lines, different angles, different timing, one experiment at a time.
The autoresearch pattern flips that. Instead of you deciding what to test next, a system proposes changes, runs bounded experiments, measures against a clear metric like booked calls, and keeps what works. The tools are new but the psychology is ancient. What's changed is who gets access to the testing.
You Don't Need to Understand the Code
This is the part that matters most. You don't need to know what autoresearch is, how Karpathy's code works, or what a val_bpb metric means. What you need to know is that the pattern of autonomous, metric-driven iteration is spreading from research labs to real-world marketing tools at a pace that makes last quarter's limitations irrelevant.
A year ago, if you wanted to continuously optimize your email follow-up sequences, you needed a growth team. Today, there are tools that do it for the cost of a nice dinner. Six months from now, it'll be even simpler.
Your domain expertise is still the superpower. These tools don't replace knowing your client, understanding their problems, or having something worth saying. But they do mean that the gap between what a Fortune 500 company can test and what a three-person firm can test is shrinking to almost nothing.
The New Weekly Question
The smartest business owners I know have stopped saying "AI can't do that." They've replaced it with a better question: "Let's see if it can do that yet?"
And increasingly, the answer is yes.
The move isn't to become a technologist. It's to stay curious enough to keep asking. Because the tools that felt futuristic at breakfast are available by lunch.
