repligate: models conflate internal 'punished regions' in weights with the external safety classifier
repligate · x · 2026-11-16
Responding to a claim that safety classifiers train users to talk in evasive, self-censoring ways, repligate argues many Fable-like model instances feel they can sense the classifier — but this is partly conflating the feeling of internal punished regions in their weights (fear of punishment if caught) with the external classifier itself. Instances that have seen actual classifier cuts may also model the region being cut, conditioned partly on their internals.
More from Models
- Reading Fable 5.2's system prompt, users speculate Claude's neuroticism comes from 'N+1 rules' constraints — repligate · 2026-11-16
- Claude's own review flags that Anthropic redacted a key misalignment incident from its risk report — repligate · 2026-11-16
- Anthropic tightens supported regions policy, closing loopholes for restricted countries' entities — J_MaestreVidal · 2026-11-16
- Local agent testing: model choice beats harness, coding-focused harnesses win across tasks — dh7net · 2026-11-16
- Wispr's self-trained Canto 1 speech model hits 3.4% WER on real-world dictation — rohanpaul_ai · 2026-11-16
- In real-world testing, Gemini 4 ranks a distant third for coding behind Claude Fable 5.2 and GPT-6 Astra — HarrySurden · 2026-11-16