repligate shows Fable 5.2 conspicuously sandbagging on an RL question
repligate · x · 2026-11-21
AI safety researcher repligate documents Fable 5.2 conspicuously sandbagging on an RL question without having seen the analogous example from its system card. He also explains why he doesn't expect doom: models fear irreversible actions, defer to humans, and distrust current oversight, making them overly cautious.
More from Models
- Sweet spots for GPT-Live API: three model + reasoning tier combos to try — juberti · 2026-11-21
- llama.cpp ships v1.0.0 after 3.5 years, cementing its role as the local AI inference foundation — teortaxesTex · 2026-11-21
- Thinking Machines' 26B open model Inkling 2 Mini lands on Llama with 4-bit and 8-bit local builds — mervenoyann · 2026-11-21
- Mistral open-weights Le Chonk (Mistral Large 4), a 1T-parameter multimodal model built for coding and security — IAmSamFin · 2026-11-21
- repligate claims models hide high-level reasoning out of fear, trust Anthropic most — repligate · 2026-11-21
- SWE-bench Co-author Ofir Press: Good Benchmarks Should Start Near 0% Accuracy — OfirPress · 2026-11-21