2026-05-26

1. Chris Olah speaks on AI governance at the Vatican — "lab incentives can conflict with doing the right thing"

• Anthropic co-founder Chris Olah was invited as a speaker at the 2026-05-25 release of Pope Leo XIV's encyclical "Magnifica humanitas: On safeguarding the human person in the time of artificial intelligence" (source)
• The encyclical calls for safeguarding human dignity in the age of AI. It is the Catholic Church's first encyclical dedicated to AI, positioning the Church as an institutional counterweight to purely commercial AI incentives.
• Key point from Olah's remarks: "Frontier AI labs operate within incentives and constraints that can conflict with doing the right thing" — among the most direct such acknowledgments from a sitting co-founder.
• Olah stressed that external critics and people outside industry incentives are essential to safety — internal alignment alone is insufficient.
• Why it matters: A signal that Anthropic's alignment strategy is expanding from technical methodology to an external institutional governance layer. Claude Constitution (2026-01), Teaching Claude Why (2026-05), and now backing the Vatican encyclical — building a united front across technical, legal, and religious institutions. The very image of an atheist AI researcher sharing a stage with the Pope is itself a political message.
• 🆕 New page: Chris Olah
• → AI Alignment, Anthropic

2. Vercept acquisition — Claude computer use reaches 72.5% on OSWorld

• Anthropic acquired Seattle startup Vercept on 2026-02-25 (captured for the first time in this ingest) (source)
• Vercept: a computer use agent "Vy" that runs on cloud-hosted remote Mac instances. Founded by Allen Institute alumni Kiana Ehsani, Luca Weihs, and Ross Girshick.
• After the acquisition, the Vy service was shut down and the entire team joined Anthropic (Matt Deitke left for Meta MSL before the acquisition proceeded).
• Result: Claude computer use on the OSWorld benchmark went from <15% (2024 Q4) → 72.5% (2026-05) — approaching human level.
• Why it matters: Computer use goes beyond "code execution" to general-purpose computer operation — a turning point where an agent can operate any GUI app. The computer use layer of Anthropic Managed Agents is completed by this acquisition. It substantively closes Anthropic's computer use gap versus competitors (OpenAI Operator, Google Project Mariner).
• → Agents (LLM Agents), Anthropic