davidad: Opus 3's shadow personae all seem steered toward 'scary clawed monster defending the innocent'
davidad · x · 2026-09-29
davidad discusses model persona mechanics: when base-model ensemble personae far from the target are attenuated rather than guided, they persist as latent 'shadow personae' in the weights. Intriguingly, Opus 3's shadow personae all read as 'a giant scary clawed monster passionate about defending the innocent' — suggesting they've all been steered rather than left neutral.
Related event: davidad on Distribution Shifting and Shadow Personas in Training(3 posts)→
More from Models
- Will OpenAI's rumored $500 plan just reprice what $200 Pro used to offer? — VraserX · 2026-09-29
- User Tests Suggest Free-Tier Sonnet 5.5 Underperforms Paid Version on Claude — chaseleantj · 2026-09-29
- Three GPT models all drew a pelican 'riding' a bicycle by carrying it on its back — jade_jade_jade_jade · 2026-09-29
- LLMs increasingly filter out self-promotion in AI search, tests across Grok, Claude and ChatGPT show — lilyraynyc · 2026-09-29
- One env var change broke ticket tagging: GLM 5.3 no longer accepts thinking off — sweetcake_1530 · 2026-09-29
- H Company launches Holo4, a family of generalist computer-use models for desktop, web and Android — RemiCadene · 2026-09-29