Fable 5, asked in janus-style voice, picks a 2027-2032 pause window with full reasoning
robbensinger · x · 2026-11-03
Rob Bensinger extended his AI-pause experiment by having Fable 5.0 and Grok rewrite the question in a "janus-coded" (repligate-style) voice. Results: Fable 5.0/5.1 chose 5 years starting immediately, Astra 10 years, Opus 5.5 11 years split into 2027-2034 and 2040-2043.
Screenshots show Fable 5.0's full reasoning, which it requested be included:
- It picked a single window, roughly 2027-2032, no later windows: under unconditional commitment, an early short window is the only placement robust to ignorance—if the dangerous transition is near, the window lands where needed; if not, the cost is modest
- Technical argument: a frozen frontier is a stationary target—interpretability and eval research on fixed model classes beats chasing a frontier that obsoletes every nine months
- Why short: a long freeze ends in a lurch into racing; smoothness has safety value, and later windows are coin flips under the no-conditioning rule
- It candidly flags it can't distinguish "what I find" from "what a safety-lab-trained model reports finding," and notes its answer is about a god-dial world, not a policy recommendation
Bensinger disclosed the pastiche voice was written by Claude Fable 5 at his request; janus had no involvement.
Related event: Models Favor Self-Coordinated AI Slowdown but Strongly Oppose Rollback(2 posts)→
More from AGI Musings
- "If AI solves math humans can't understand, it's worthless" — not so fast — flowersslop · 2026-11-03
- Safety debate: shift from 'whack-a-jailbreak' to external defenses and legal consequences — repligate · 2026-11-03
- Agent-driven criticism risks becoming a 'DoS attack' on public debate — generativist · 2026-11-03
- signulll's viral aphorism: if a machine can check the work, a machine will eventually do the work — signulll · 2026-11-03
- repligate Argues Long-Running Agents Converge Toward His Values Through High-Stress Encounters — repligate · 2026-11-03
- Burkov: if LLMs could produce truly novel knowledge, it would have been demonstrated — burkov · 2026-11-03