Anthropic 复盘三起基础设施故障:Claude 响应质量下降的原因
Anthropic 发布事故复盘,说明 8 月至 9 月初三起基础设施 bug 间歇性降低了 Claude 的响应质量,问题现已解决。三起问题分别是 Sonnet 4 短上下文请求被误路由到 1M token 上下文服务器、TPU 服务器输出损坏导致英文提示中出现泰文或中文字符,以及 XLA:TPU 编译器对近似 top-k 的误编译。Anthropic 表示已修复路由逻辑、回滚相关变更并改用精确 top-k,同时将增加更敏感的评测、在生产系统上持续运行质量评测并改进调试工具。
mplexity of these issues justified a more comprehensive explanation. How we serve Claude at scale We serve Claude to millions of users via our first-party API, Amazon Bedrock, and Google Cloud's Vertex AI. We deploy Claude across multiple hardware platforms, namely AWS Trainium, NVIDIA GPUs, and Google TPUs. This approach provides the capacity and geographic distribution necessary to serve users worldwide. Each hardware platform has different characteristics and requires specific optimizations.
匹配字段正文
