Anthropic published a piece called “When AI builds itself,” and one line stopped me: more than 80% of the code merged into their own systems is now written by Claude — up from low single digits when Claude Code launched in early 2025. Their engineers ship roughly eight times the code per quarter they used to. And on the hardest, most open-ended engineering problems, Claude's success rate has climbed past 76%, a fifty-point jump in six months.
I read those numbers and felt a small jolt of recognition — not surprise. Because I'd been feeling this for months without a word for it. I wrote here in May that my thought-to-execution speed had collapsed from weeks into hours. I wrote about watching Claude connect the small tools I'd built over years into something I couldn't have assembled myself. What I was describing, in a leader's plain language, the frontier was now describing in a curve. The number is just my hunch, measured.
The phrase they reach for is recursive self-improvement — an AI system designing and building its own successor, with less and less of us in the loop. They sketch it as stages. First, AI accelerates the people doing the work. Then development becomes substantially automated while humans still set direction — a hundred people doing the work of a hundred thousand. Then, in the last stage, the system builds the next system, and the next, on its own.
What's striking is who is saying it. This is the lab itself — the people closest to the curve — saying this could arrive sooner than most institutions are prepared for, and floating something they call a global coordination mechanism: a way to slow, even pause, that only works if everyone near the frontier stops together, under rules outsiders can verify. A company asking, out loud, for the option of a brake on its own engine. That is not a sentence I expected to read this year.
I sit in an odd seat for this. I'm not a frontier researcher; I'm a leader who went through an AI transformation by having weeks of heart-to-heart conversations with my team, not by reading scaling laws. So the report reaches me twice. Once as exhilaration — the compounding I've been living is real, named, and bigger than my one office. And once as a quiet unease — the same loop that made me faster is the loop that, followed to its end, stops needing the hands that started it.
I keep landing on the human question underneath the technical one. If a system can build its own successor, then the things that don't automate — how we secure it, how we watch it, what we point it at — matter more, not less. The judgement, the intent, the why. The part of my own work AI hasn't touched is exactly the part I once thought was the soft, optional layer: the conversation, the direction, the care about who it's for. Turns out that was never the soft layer. It was the load-bearing one.
I'm not going to pretend I know whether the global pause is wise or even possible. I don't. What I notice is the honesty of a lab naming the thing while it's still a slope and not yet a cliff — saying it plainly enough that even someone in my chair, far from the lab, can feel where it's heading.
So I'm sitting with the double feeling and not resolving it today. My daughter will grow up on the far side of this curve. My team is already living on its near side. And the most useful thing I can do, I think, is the unglamorous thing the report quietly implies: stay in the loop on purpose, for as long as staying in the loop still means something — keep my hand on the why, even as the how writes itself.
Here's the question in its plainest form. When AI can do the doing — and all the things we called intellectual — what's left for humans to do?
Maybe the answer is smaller and stranger than I expected: navigating expectation is the skill. Knowing what to ask for, what's reasonable, what good looks like, when to push and when to accept. Not everyone will know how to navigate AI — the same way not everyone knows how to build an operating system. And honestly, is everyone even a pro at operating a phone? The tool got infinitely capable; most of us still use a sliver of it.
So perhaps the new literacy isn't doing the work. It's holding the right expectations of a thing that can do almost any work — and that, it turns out, is not evenly distributed at all.