A few months ago I took my sites off WordPress. Not as a project with a timeline and a launch date, but in an afternoon. I described what I wanted, Claude turned it into code, and by the end of the day something was running that I had failed to get done in weeks before.
I have written about that already. What stayed with me was not the move itself but the feeling of that afternoon. The sense that the ratio between what I wanted and what I could actually make had quietly tipped.
This week Anthropic put a number on that feeling.
The numbers
In a thread about what they call recursive self-improvement, AI helping to build AI, Anthropic shared internal data. The core of it: their own engineers now ship on average eight times as much code per quarter as they did in the years before.
It was not only volume. On open-ended coding problems, the kind where the answer is not fixed in advance, Claude now reaches a success rate of 76 percent. Six months earlier that figure was fifty points lower. Many of their engineers say the code quality is now on par with human work, and they expect it to be better within a year.
One test stayed with me longer than the rest. Give the model code that trains a small AI model, and ask it to make that code faster. A skilled engineer needs four to eight hours to reach a fourfold speedup. In May 2024 their model averaged roughly three times. This spring the newest model reached a factor of 52.
Anthropic is honest about the edge of what this means. They write themselves that none of it guarantees AI will start improving itself, and that it stays unclear whether the model has research judgment: the ability to choose the right problems. Hold on to that sentence. I will come back to it.
The mirror
I am not a frontier lab. I am one person with a few sites, a newsletter, and a pile of ideas that used to sit untouched because execution was too expensive.
And still I recognize the graph. Not the number, that one is theirs, but the shape. The hundred small handoffs that filled a day, the copying, the switching between tabs, the formatting, are mostly gone. What I used to postpone because it was too much hassle, I now do over lunch. My own eight times does not live in smarter thoughts. It lives in the fact that the thoughts get out the door faster.
And there is something else, something to do with me. I am creative and headstrong. Concepts come to me easily and fast, and I see a thing before a single word is down. Typing was the bottleneck for years, sitting between that picture and anything that existed. I am still typing this, but what I like most, and do most, is putting it in with my voice. Voice First. I will tell you about that another time. It is the same move as that eight times. Not that I think better, but that less sits between the thought and the output.
That is exactly where I get careful.
What the number does not measure
Three things I hold up to the light before I go along with the story.
First: Anthropic is grading its own homework. This is internal data, shared by the company with the most to gain from precisely this story. That does not make the numbers false, but it does not make them an independent measurement either. A company that sells an accelerator and then measures how fast it accelerates is telling a story in which it plays the lead.
Second, and this is the real one: eight times the code is not eight times the value. Code is the easy thing to count. It is volume, and volume is exactly what a machine is good at scaling up. But value rarely lives in the quantity. It lives on the last mile, where you decide which problem is worth solving at all. Anthropic says as much itself, in that one sentence I asked you to remember. Their model improves the next step of a researcher 64 percent of the time. Impressive. But choosing the step is not the same as choosing the direction.
The third follows from that. More output is not more thinking. In an earlier piece I named the difference between the hauling and the thinking. The acceleration I feel, and the one Anthropic measures, sits almost entirely in the hauling. The moving, the building, the executing. The thinking, deciding whether a thing is worth doing, has not become eight times faster. Not for me, and judging by their own caveat, not for them.
What I take from it
The temptation with a number like this is to extend the line. Eight times this year, so imagine next year. But a line that climbs on the axis you can measure easily says nothing about the axis you cannot measure.
I use the acceleration every day, and I would not go back. But I try to keep asking myself one question. Not how much I get out the door, because that answer is easy now. Whether I still get the right things out the door. That is the measure missing from the graph, and over time it is the only one that matters.
Being able to make eight times as much is a gift. It only turns into a problem when you forget that producing faster and choosing better are not the same thing, and that only the second is something a machine will not take off your hands any time soon.
This piece belongs to an ongoing line on AI as an operator. Read also how I left WordPress behind and the assistant that executes.