In a blog post titled "When AI builds itself," Anthropic researchers Marina Favaro and Jack Clark argued the world should have the option to slow or temporarily frontier AI development, citing the risk that models could reach "recursive self-improvement." The post reports that, as of May 2026, Claude authored more than 80% of the code merged into Anthropic's codebase, and that the typical engineer now merges roughly 8x as much code per day as in 2024. Clark told BBC Newsnight, "You want the option to be able to take your foot off the gas and put your foot on the brake," and said fully self-written code could be possible within two years. Anthropic frames any as conditional: it would slow down only if other frontier labs did so verifiably, and says The Anthropic Institute will research the verification systems such coordination would require. The appeal lands days after Anthropic confidentially filed for an IPO that could exceed a $1 trillion valuation.
What happened
In a blog post titled "When AI builds itself," published June 4, Anthropic researchers Marina Favaro and Jack Clark argued that the world should have the option to slow or temporarily frontier AI development so that "societal structures and alignment research" can keep up. They tie the argument to "recursive self-improvement," a threshold at which an AI system could autonomously design and develop its own successor. Anthropic states this has not happened and is "not inevitable," but warns it "could come sooner than most institutions are prepared for."
The evidence Anthropic disclosed
The post is built on internal data, not only rhetoric. Anthropic reports that, as of May 2026, more than 80% of the code merged into its codebase was authored by Claude, up from low single digits before Claude Code launched in February 2025. It says the typical engineer now merges roughly 8x as much code per day as in 2024, and cites outside benchmarks (SWE-bench saturation, and METR's finding that AI task length is doubling about every four months) as corroboration. In one example, Anthropic says Claude shipped over 800 fixes in April 2026 that cut a class of API errors a thousandfold.
What Clark told the BBC
In a BBC Newsnight interview, Jack Clark said, "You want the option to be able to take your foot off the gas and put your foot on the brake." He told the BBC that Claude is now operating on code of which roughly 80% the system wrote itself, and that reaching 100% "is possible within two years."
The is conditional, not unilateral
Despite some headline framing, Anthropic did not call for an immediate one-sided halt. The post states a would require "multiple well-resourced labs at or near the frontier, in multiple countries, agreeing to stop under the same conditions," plus the ability to verify that others have actually stopped. Anthropic says it would slow or only if peers did so verifiably, and that The Anthropic Institute will research the verification systems (audits, provenance tracking, attestation of compute) such coordination would need. The company notes major AI firms have not committed to any .
Reception and why it matters
The appeal arrives days after Anthropic confidentially filed an S-1 with the U.S. Securities and Exchange Commission on June 1, ahead of an IPO that reporting says could exceed a $1 trillion valuation; its latest private round valued the company near $965 billion. That timing drew skepticism, with some commentators questioning why a lab heading for a trillion-dollar listing would urge hitting the brakes, and critics framing the post as competitive positioning. Others counter that detailed internal disclosure of AI-driven research acceleration is itself a substantive contribution to the governance debate.
For practitioners, what to watch
Industry-pattern observation: making any slowdown operational rather than rhetorical would require concrete machinery, so useful signals include published cross-lab verification protocols, joint safety exercises, adoption of provenance and compute-attestation tooling, and any government motions on moratoria or mandatory safety testing. The nearer-term signal for engineers is the automation curve itself: if code authorship and experiment-running keep shifting to models, human review and research taste become the bottlenecks Anthropic says it is already encountering.
Scoring Rationale #
A top-three frontier lab disclosed substantial internal evidence that AI now drives a majority of its own engineering (over 80% of merged code) and publicly argued for the option of a verifiable, coordinated slowdown, an agenda-setting intervention in the recursive self-improvement debate. Coverage was broad (BBC, CNBC, Decrypt, SiliconANGLE) and the timing alongside a potential $1 trillion-plus IPO amplified its policy and market salience. It remains an appeal rather than a binding action or a model release, which holds it just below the industry-shaking tier.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.