用户吐槽 Opus 被 RLHF 训出"各打五十大板"的固执习惯
teortaxesTex · x · 2026-11-11
tracewoodgrains highlighted Claude Opus's verdict on a debate—"Alice is winning on the argument, Ben is winning on temperament"—as evidence of a persistent RLHF-instilled tic: the split-the-baby, say-something-nice-about-everyone approach, one of Opus 5.5's most stubborn habits. teortaxesTex replies that a system prompt easily defeats it, but he refuses to use one: he wants to win so devastatingly that Claude admits he "eviscerated the other guy" despite its conditioning.
「Fun」频道最新
- 梗图：Python 程序「说流利 neuralese」震惊神经网络 — nabla_theta · 2026-11-11
- AI 生成恶搞海报:《The Spy Who GPU'd Me》 — johnmccrea · 2026-11-11
- 用户让 Claude 总结 Anthropic 封禁声明,反被指"传播虚假信息" — sethlazar · 2026-11-11
- 论文未放代码，网友用 Claude 2小时28分复现，20天冒出237个repo — ChongZzZhang · 2026-11-11
- OpenAI 工程师暗讽 Codex 听写功能：'没听说过，肯定是垃圾' — chrisalbon · 2026-11-11
- 网友调侃：让 Devin 接管 Semrush 和 Search Console,SEO 从业者自愧不如 — AAAzzam · 2026-11-11