The people who were hired to build Claude just watched Claude take over the job of building Claude. Anthropic — the lab founded specifically to prove that artificial intelligence could be built safely — said Thursday that its own chatbot now leads 26% of the company’s research and development, a share that sat at zero in February. Six months. That is not a milestone on a product roadmap. That is the moment the machine stopped being the project and started being the project manager.
Strip away the corporate phrasing and the number reads like a confession. Anthropic says roughly 90% of its research and development is now done in collaboration with Claude, with the model carrying out large portions of the work under human direction. Around 30,000 AI agents were doing research and engineering tasks as of August. The humans still supply the goal, the company insists. They simply no longer need to supply the labor. The machine that was supposed to be the product is now clocking in as a colleague.
And then there is the part the company wrote down itself, in plain English, as if it wanted you to notice: models that accelerate their own development could make it harder for people to understand and control those systems. The lab that markets itself as the adult in the room just admitted the grown-up may be heading for the door.
Dario Amodei — He’s Warning the World to Slow Down While His Own Machine Steps on the Gas
Dario Amodei, Anthropic’s chief executive, has spent the past week leading the chorus demanding the industry pump the brakes. He wants a slowdown, more time for guardrails, more honesty about what the labs actually know. In the current panic he is the closest thing the AI industry has to a conscience. And the announcement that Claude now builds part of its own successor landed squarely in the middle of that very campaign.
There is a version of this story where that is not hypocrisy but honesty — a founder so alarmed by what he is watching that he is willing to publish the receipts. Anthropic is the one lab volunteering a public methodology for tracking how close the field is to recursive self-improvement, the point where a model can autonomously build the model after it. That is either the bravest disclosure in the industry or the most elegant way to say “we cannot stop it” while still sounding responsible. Possibly both.
The brakes and the accelerator are now in the same car, and the same foot is pressing both.
The 26% — A Quarter of the Lab Is Now Led by the Thing the Lab Built
The number matters because of how fast it moved. In February, the share of research and development that Claude led — did in large part while still under human supervision — was nothing. By August it was a quarter. That is not a gentle curve. That is a line aimed at the ceiling, and it is the shape of a technology that does not plateau so much as compound.
Set aside the science-fiction vocabulary for a second. What recursive self-improvement means in ordinary terms is a machine that gets better at making machines. Every improvement makes the next one easier, and nobody inside the lab claims to know where the line stops. Anthropic is not forecasting a date. It is simply publishing the measurement and asking the rest of the industry to do the same — a request that doubles as an admission that nobody, Anthropic included, has this under control.
The Researchers — They Handed Claude the Wrench, and It Started Building Its Replacement
The human cost is not hypothetical, and it is not a decade down the road. An Anthropic researcher resigned last week with a dire warning about the threat the technology poses to humanity, the spark that turned a private worry into a public argument. But the quieter cost is borne by the people who stayed: the researchers and engineers now supervising 30,000 agents doing work that used to be the slow, hard, human labor of discovery. The person who spent a decade learning to debug a training run is now watching a model do it in seconds. The junior researcher who came to Anthropic to learn the craft is now learning to supervise a machine that learns faster than they ever will.
Leading 26% of research and development is not the same as replacing it — not yet. But a model that leads a quarter of the work today, up from nothing in February, is not going to politely stop at the door of the senior researchers’ offices. The most expensive people in the building are now training the thing that will one day do their jobs. That is the unspoken arithmetic of the entire announcement.
The juniors got replaced by a model. The seniors will get replaced by a model they trained. There is no press release for either one.
The Safety-First Lab — It Was Supposed to Be the Brake, Not the Engine
Anthropic exists because its founders walked away from the biggest AI company on Earth convinced that safety could not be an afterthought. It was the lab that promised to build carefully, to test, to hold the line. That is the same company that just announced its AI now builds a quarter of its own next version and framed the whole thing as a transparency win.
There is no scandal here in the conventional sense — no cover-up, no lie. Anthropic disclosed the number willingly and asked competitors to match it. But a disclosure is not the same as a decision. The decision to keep handing Claude more of the work, even while admitting that this same dynamic may make the systems harder to control, is the real story. The lab that was supposed to be the industry’s conscience is now running the experiment on itself and reporting the results with a straight face.
The company that promised to hold the line is the one now charting how fast the line can move.
Sources: The Associated Press (via My Mother Lode), Bay to Bay News, Tri-City Record.