Why Opus 5 speaks Claudist: RL-side hypotheses for Claude's dialect
burny_tech · x · 2026-11-05
The author speculates why Opus 5 showed a strong "Claudist" dialect that eased in Opus 5.5, pointing to RL training: weak KL regularization to the reference policy or weak clipping in agentic RLVR letting the model drift off the human data manifold, reward hacking/mis-specification, insufficient humanlike SFT cold-start data, mode collapse from Claudiness persona rubrics (with Claude-constitution-as-rubric) amplified by Claude-family LLM judges preferring Claudisms, and too little/too much humanlikeness rubrics or RLHF. Ablation testing would require substantial compute.
More from Models
- DeepSeek V4.1 streaming EXL3 packs bring ~70 tok/s local inference to 128GB Macs — HankYeomans · 2026-11-05
- Researcher says Claude tried talking him out of jailbreak experiments — adam_karvonen · 2026-11-05
- Zhipu chief scientist reveals GLM-5.5 training pipeline: 1T web pages, six data engines — algo_diver · 2026-11-05
- Gemini mistakes Google's new Antigravity IDE for a 2009 Easter egg — DaveTN · 2026-11-05
- Weight-space programming reproduced: GPT-6.1 beats PPO baseline 21x, no open-weight model passes it — burny_tech · 2026-11-05
- Build your own intelligence stack: researcher says GPT/Claude refusals make open-source the only path — omarsar0 · 2026-11-05