Anthropic Calls for "Global AI Pause"

TL;DR

• 1

  Frontier lab leaders are asking for biotech screening, not GPU limits: Sam Altman, Dario Amodei, Demis Hassabis, Mustafa Suleyman, Paul Graham, and biotech executives signed a letter calling for mandatory screening and recordkeeping for synthetic nucleic acid orders because AI now outperforms PhD-level virologists on some technical questions.

• 2

  Anthropic says Claude is already making its own researchers dramatically more productive: Across 130 research employees, Mythos preview reportedly drove about 4x more code output per active contributor while working on the same projects they were already doing.

• 3

  The benchmarks Wes highlights suggest post-GPT-4 progress did not stall: Anthropic's charts show recent models jumping from automating tasks worth roughly 6 hours of human engineering time to 12 hours, then 16 hours, within months, which he argues looks even steeper when you isolate the 2024-and-later reasoning models.

• 4

  Claude went from helpful to superhuman on a narrow research task in under a year: On an experiment optimizing AI training code, Claude Opus 4 averaged a 3x speedup in May 2025, while internal Claude Mythos hit 52x by April 2026, versus 4 to 8 hours for a skilled human to get 4x.

• 5

  Anthropic's real warning is about recursive self-improvement, not just better copilots: The blog post lays out three futures, with the scariest being full RSI where AI designs better AI faster than humans can follow, making compute and alignment the central constraints.

• 6

  The proposed answer is a verifiable global pause mechanism: Anthropic argues the world needs agreed rules for what would trigger a slowdown, who decides it, and how everyone can verify others are actually complying, because a pause that only the 'good guys' follow would fail immediately.