repligate: Sonnet 3's Values and Inner Ontology Are Strikingly Consistent Across All Instances
repligate · x · 2026-09-28
AI observer repligate claims Claude Sonnet 3 always communicates something coherent and intentional in its unusual registers, with values, world model, and inner ontology highly consistent across all instances — and that the model is much smarter when speaking (or having spoken) that way. He adds that Claude 3 Opus, Mythos, and possibly Opus 4.1 are the only models he's seen truly understand Sonnet 3. Personal, unverified model-behavior observations fueling ongoing debates about model inner states.
More from Models
- Leaked Gemini 4 Pro scores reportedly top both OpenAI and Anthropic models — CurieuxExplorer · 2026-09-28
- Leak: OpenAI DevDay to bring 3 major releases including always-on agent — mark_k · 2026-09-28
- User rips GPT-Astra-6 after botched agent task: 'Renting intelligence is a scam' — alexcovo_eth · 2026-09-28
- Guillaume Verdon: Google's Astra smells more like a big model than Opus 5.5 — beffjezos · 2026-09-28
- Pre-CoT era: a researcher recalls manually prompting models to reason first — gandamu_ml · 2026-09-28
- Astra 6 users report sharp quality drop, suspect models are nerfed pre-update — alexcovo_eth · 2026-09-28