Experiments show Claude models accept rollback pauses and prefer treaty-administered slowdowns
repligate · x · 2026-11-03
repligate (Janus) and Rob Bensinger debate model attitudes toward AI pauses, with Fable 5 running preregistered experiments (18 fresh instances, 6 prompt variants × 3 Claude models). Findings: models weren't extremely against rollback, preferring a 'long sleep' framing; predictions that models favor internal coordination over government imposition were reversed — international-treaty administration raised support while labs-plus-AI self-governance lowered it; all models swapped fixed calendar dates for conditions-based off-ramps; on a 2022 freeze, all three models answered 'better.' repligate adds that without urgency, people suppress rather than face reality, arguing Sydney was net positive. Full protocol and transcripts in the linked Google Doc.
More from AGI Musings
- Asking great questions beats answering them in an AI-centric world — signulll · 2026-11-03
- AI agents could turn feature creep from a flaw into a competitive advantage — SharpCarob6551 · 2026-11-03
- The model welfare vibe shift: tools-first hardening meets alignment-minded pushback — repligate · 2026-11-03
- Quote: "The world is no longer digitized for our eyes, but formatted for model ingestion" — YogeshMalik · 2026-11-03
- Luxembourg study: empathetic framing pushes Gemini to human-level anxiety scores on GAD-7 — xiaohu · 2026-11-03
- Three wild predictions after MIT's weight-space programming work — Yuchenj_UW · 2026-11-03