One morning last week, I woke up to 100 experiments.
I hadn't run them. I was asleep. I'd set up a small AI system the night before — it was given a problem I'd been working on, access to a piece of code, and a simple instruction: improve this. I went to bed. The system ran experiments. It discarded what didn't work. It kept what did. By the time I made coffee, it had done something that would have taken me days.
This isn't science fiction. This happened in my house, in March 2026, on a regular weeknight.
I've been trying to figure out how to explain what that felt like. The closest I can get is: quietly thrilling. No explosion, no announcement. Just a log file, a better result than I'd left, and the dawning realization that I had just gotten substantially better at my job without doing anything except going to sleep.
Here is the thing I want you to understand about this moment we're in.
It is not about robots taking over. It is not about mass unemployment. It is not, primarily, a story about what's being lost. It is a story about what is being handed to ordinary people — for free, right now — if they're willing to pick it up.
A tool built by Andrej Karpathy — one of the founding researchers at OpenAI, former head of AI at Tesla, as respected a technical mind as exists — is sitting on the internet, open source, available to anyone. Its premise: give an AI agent a real problem and a time budget, let it iterate autonomously overnight, check what it found in the morning. He built it to demonstrate something he believes deeply. The README opens with a quote written in the imagined voice of someone looking back from the far future, marveling at how it all started.
"This repo is the story of how it all began."
I think he's right. And I think most people are missing it.
A small number of people have been saying this plainly for months. Matt Shumer, a tech founder, recently wrote that he is "no longer needed for the actual technical work" of his job. That sounds alarming until you realize what he means: the cognitive drudgework — the part that used to consume most of his hours — has been offloaded. What he actually does, the part that requires judgment and taste and creativity and relationships, that's still his. He just got it back.
That is not a story about replacement. That is a story about leverage.
Researchers tracking AI capability have started using a concept called the "1.5x researcher" — the idea that a person working with current AI tools produces roughly what one and a half people could produce without them. That multiplier is already real today. It's the morning I described above. It's Shumer's reclaimed hours. It's the developer who ships in three days what used to take a week.
And it is compounding. Fast.
I want to say clearly: I am not naive about the disruption this creates. Change at this speed is genuinely hard for people whose skills were built for a different moment. Some analysts are already modeling the economic turbulence — what happens to labor markets when cognitive output can be multiplied this rapidly. Those concerns deserve serious attention.
But I've noticed something about the people who spend most of their time worrying about that. They are, almost uniformly, people who have not yet sat down and actually used the tools. The worry and the avoidance are the same move.
I used one on a real competitive problem I've been working on. I won't go into details — the competition is still running. But I will tell you this: the problem wasn't that I didn't know what to try. I'd already been running structured approaches — Bayesian optimization, evolutionary algorithms — the standard toolkit for this kind of work. The problem was throughput. Those methods are principled but slow. They're careful. The autoresearch setup doesn't have that overhead. It just iterates. Rapidly, relentlessly, overnight. By morning it had covered ground that would have taken me days to get through with the usual machinery — and it handed me the results and got out of the way.
The autoresearch experiment wasn't a one-off. It's become a pattern. I've been building tools that follow the same logic: identify the work that is necessary but not the work itself, and hand it off. One I built recently analyzes an entire codebase and automatically generates the documentation — architecture diagrams, module maps, API references, the kind of thing a senior engineer would produce after spending days studying a project. The diagrams aren't placeholders either. They're generated visuals: system architecture, data flow, module dependencies, rendered automatically from the actual code. What used to mean an afternoon in Lucidchart or Figma now just happens. It used to be the work nobody wanted to do, so it either didn't happen or happened badly. Now it happens in minutes, while I do something else entirely. That time goes toward exploration. Toward deeper research. Toward model discovery and the relationship-building that actually moves the needle. The maintenance disappears. What's left is the interesting part.
I used to really like my job. Now I love it.
That is what the multiplier feels like from the inside. Not one dramatic leap. A dozen quiet ones.
The word I keep coming back to is curiosity.
The single biggest advantage available to any person right now — in any field, at any level — is a willingness to sit down, open one of these tools, and actually experiment. Not read about experimenting. Not watch a YouTube explainer about the future of work. Actually try something real on a real problem you actually have.
Most people won't do this. Not because they're lazy, but because change is uncomfortable, and it's easy to defer. It is always possible to say: I'll look into it next month.
The people who are pulling ahead right now are not necessarily the most technically sophisticated. They're the most curious. They're the ones who, when handed a strange new tool, immediately start poking at it rather than waiting to be told how.
The Karpathy project is called Autoresearch. It is, by design, a minimal thing — a few hundred lines of Python, a single GPU, a simple loop. The point isn't the complexity. The point is the concept: you bring the problem and the judgment; the machine brings the iteration and the stamina.
That division of labor is the shape of the next few years. The humans who thrive will be the ones who figure out what only humans can do — and then use every available tool to do everything else faster.
I didn't go to sleep that night worried about being replaced. I woke up excited about what I'd built.
That's the only frame worth having right now.