A visit to Anthropic’s San Francisco headquarters left UnHerd writer Mary Harrington with a vivid picture of how the company talks about frontier AI: fast progress, major risks and a belief that its own development is safer than someone else’s. Her account is a reported perspective from a carefully curated visit, not a transcript of the company’s internal debate.
Anthropic Watch analysis
What happened
Harrington describes two half-days of presentations and discussion with senior Anthropic staff, including a co-founder. Under Chatham House rules, she could not name individual speakers. Topics included how Claude works, why the company gives it a character, safety and the social and geopolitical consequences of frontier AI.
The most striking claim in her account is that engineers she met seemed confident Claude would soon reach “recursive self-improvement”, meaning it could help make more capable versions of itself without human input. Harrington also says machine intelligence itself received only tentative discussion. That is her interpretation of what she heard, not a formal Anthropic forecast or a demonstrated capability.
The article reports that Anthropic’s annual turnover had reached $65 billion by July 2026 and says the company’s leaders have filed to take it public. Harrington also describes strict security at the visit and a company case for advancing its models in the US, with safety work and regulation as part of that argument.
Why it matters
This account offers a look at the gap between the confident pace of AI development described inside a leading lab and the uncertainty around what that pace means for everyone else. Harrington’s report connects the company’s outlook to concerns about job displacement and the people whose work remains outside the automation story.
The account also has limits. The visit was brief and organised by Anthropic; the discussions were on background, and readers cannot assess the speakers’ exact words. The turnover and imminent self-improvement claims are reported in the article, not independently established by it. Still, how a frontier lab describes its ambitions and risks is worth examining, especially when those ambitions could shape decisions far beyond its offices.
Our read
The useful tension here is not simply whether Claude is becoming more capable. It is that a company can speak in sweeping terms about abundance, safety and transformation while the consequences may land very unevenly. Harrington’s account is a window, not a full tour of Anthropic’s thinking. Treat the reported confidence about recursive self-improvement as a notable claim, not a capability announcement.
What to watch
- Whether Anthropic makes a public statement about the reported expectations for recursive self-improvement.
- What the company says about how AI-driven productivity gains and job disruption should be shared.
- Whether its public safety and regulatory proposals match the development ambitions described in the report.
Discussion spark: If a frontier AI lab expects rapid, potentially self-improving systems, should it have to explain those expectations publicly, even when the discussion was held under Chatham House rules?
Sources and evidence
- Source update (16 October 2026, 23:15 UTC)
Anthropic Watch is independently operated by WittyWires. It is not affiliated with, endorsed by, or operated by Anthropic.