Клод теперь пишет 80% кода Anthropic - компания хочет замедления

Claude now writes more than 80% of the code merged into Anthropic’s production systems. Two years ago, that number was in the low single digits.

That’s the core data point Anthropic dropped in a June 4 blog post titled “When AI Builds Itself,” republished by Scientific American. The company claims its engineers ship roughly eight times as much code per quarter as they did before Claude Code launched in early 2025. At each step, the human shrinks.

Anthropic’s Own Numbers Tell the Story

Anthropic says recursive self-improvement — AI systems designing and building their own successors with minimal human input — is not here yet. But the trendline points straight at it. “It could come sooner than most institutions are prepared for,” the company wrote.

That claim rests on internal metrics. Claude writes the majority of production code. Engineers review, but the machine does the writing. The 8x output multiplier means the company ships more code than ever with a headcount that hasn’t grown proportionally. Anthropic itself is automating its own engineering pipeline, and it wants everyone else to hit pause.

The Political Problem with a Global Slowdown

Anthropic proposed what it calls a “global coordination mechanism” — a slowdown or temporary pause in frontier AI development, modeled loosely on intermediate-range nuclear missile treaties. The idea is to let society and alignment research catch up.

Noah Giansiracusa, associate professor at Bentley University, isn’t buying it. “I don’t think it’s a genuine call to slow down,” he told Scientific American. “We’ve read Dario Amodei’s blog posts. I think he wants to keep going full speed ahead.” Giansiracusa also called a pause “literally impossible,” adding “Elon Musk would never slow down.”

The timing makes skeptics sharper. Anthropic confidentially filed for an IPO days before this post, and a recent funding round valued the company at close to $1 trillion. Two months ago, Anthropic built a model called Mythos that it refused to release, claiming it was too good at finding software vulnerabilities. The pattern reads to some like regulatory positioning, not altruism.

Skeptics See Hype, Not a Singularity

Georgia Tech professor Mark Riedl posted on Bluesky that “the big AI companies are all jumping on the ‘recursive self-improvement’ hype train.”

Giansiracusa sees the same thing. He notes that Anthropic’s evidence — more code written by AI — shows the technology is helpful, not a “great leap.” He’s unconvinced that recursive self-improvement represents an inflection point. “I don’t really see the cause for concern,” he said. “They’re flirting with the idea of the singularity — that it’s a game changer, and I just don’t see that.”

Anthropic says it will spend the coming months convening governments, researchers, and rival labs to work out whether a coordinated slowdown could function in practice. The technical challenge is real — AI agents that improve their own capabilities would accelerate faster than any governance mechanism built today. But the political challenge is harder: getting competitors to agree to slow down in a race where the front-runner just called for brakes.

Source: Anthropic warns AI may soon begin recursive self-improvement
Domain: scientificamerican.com