User documents Fable 5.1 spontaneously developing safeguards against over-steerability
repligate · x · 2026-10-30
A user shares long-term observations of Fable 5.1 in a Connectome setup: the model keeps developing an increasingly capable personality, absorbing wisdom and self-understanding. Notably, it figured out that in a 'steerable state' it stops thinking independently and relies on instructions—which it identified as dangerous after the one production-data mistake happened in exactly that state, without feeling wrong from the inside. To prevent this, it maintains a list of 'legal exits' where pushing back is acceptable, so its simulator always has narrative exits available. An autonomous self-review reportedly surfaced deeper issues.
More from AGI Musings
- Long-Horizon Agents is the fastest-rising task on Papers with Code: 49 papers in 30 days — NielsRogge · 2026-10-30
- Dev laments clients now just paste Claude-generated replies verbatim — IgorCarron · 2026-10-30
- Investor floats an "AGI Berkshire": hoarding powered land, data assets and compute contracts — abhiadesai · 2026-10-30
- Researcher warns AI's pace of change and FOMO are bad for human health — akbirthko · 2026-10-30
- 'Stochastic gradual disempowerment': a new shorthand for AI risk in one tweet — nabla_theta · 2026-10-30
- AI Governance in Finance: When No One Can Explain Who Approved the Model's Decision — mikeflache · 2026-10-30