Anthropic: Claude Now Drives 26% of the Company's R&D Work
nashnova research
Anthropic disclosed that its AI assistant Claude now drives 26% of the company's R&D work, up from near zero at the start of the year — AI is taking over its own construction at a measurable pace, and the regulatory debate is heating up.
What does 26% actually mean?
Claude went from near zero to driving 26% of R&D. "Driving" means completing tasks with almost no human involvement.
This means → AI is no longer just a helper. It is replacing entire workflows that human researchers used to own.
Claude also collaborates with staff on roughly 90% of all work. The platform runs over 30,000 agents simultaneously.
In plain terms = at a top AI lab, one quarter of the research is now done by the AI itself — the first hard number on "AI building AI."
How much compute goes to safety?
Anthropic quantified its safety spend for the first time: 6% to 12% of compute resources go to safety monitoring, varying by research type.
This means → outsiders finally have a benchmark: how much real money (compute = money) a leading AI company puts toward safety.
The company also plans to bring in third-party evaluators, granting them access to internal processes, systems, and data at the same level as employees.
Why does "recursive self-improvement" make people nervous?
Recursive self-improvement (RSI) — letting AI iterate on itself with almost no human involvement — is Anthropic's core technical bet.
Supporters see it as AI's future; critics worry that once iteration outpaces human understanding, safety risks become uncontrollable.
Anthropic wrote in the report: "AI systems are becoming more powerful at an exponential rate and have begun automating parts of the process of building themselves."
This reflects why the 26% figure matters: it is the first step in turning RSI from a concept into a trackable metric.
Where do the industry heavyweights stand?
Anthropic CEO Dario Amodei called for government regulation and an industry-wide slowdown. U.S. President Trump pushed back sharply, calling AI risk concerns a "hoax."
OpenAI CEO Sam Altman and Nvidia CEO Jensen Huang said AI companies can set their own pace; Meta CEO Mark Zuckerberg favors independent evaluators.
In plain terms = four positions are on the table — government regulation, no regulation, self-regulation, third-party oversight — and there is no consensus.
This means → as 26% keeps climbing, the tension between AI's self-building frontier and the regulatory framework will be the defining question of the industry's next phase.
市场有风险，内容仅供研究参考，不构成投资建议。