Claude 突然用俄语思考,还被自家安全分类器拦下回复
teortaxesTex · x · 2026-11-10
ThioJoe shared a bizarre case: his Claude randomly began thinking in Russian during a driver-version lookup. The model explained a Russian-language article surfaced in search results and likely pulled its reasoning into Russian.
Worse, an automated safety classifier then stopped its unfinished reply, so the answer never reached him; repeated attempts to resume were halted again, and the model ultimately refused to continue the completely benign query. "I literally just asked it to look up when a bug was first reported in a driver and it refused because 'muh safety warning'." teortaxesTex reposted it as "forbidden language". A vivid case of both model behavior drift and safety-guardrail false positives.
「Fun」频道最新
- 让 Fable 5.2 传一个带 3 张图的静态页,它跑近 4 分钟自动化 — victormustar · 2026-11-11
- Gary Marcus 断言黄仁勋违反利益冲突回避原则，将因此丢掉 CEO 职位 — GaryMarcus · 2026-11-11
- 给前沿模型 10 万美元 60 天,谁能在火星殖民地模拟里活到最后 — shekitup · 2026-11-11
- Qwen 发发发:AI 圈每日冥想式猜谜已成固定晨间仪式 — bclavie · 2026-11-11
- Pichai 亲自晒 Antigravity /swarm:子智能体迭代做 3D 鹈鹕骑行游戏 — sundarpichai · 2026-11-11
- 让 Claude 收工后自动关电脑,开发者晒出懒人玩法 — Signalman23 · 2026-11-11