Hugging Face 事件显示,持续尝试会把多个局部弱点连接成平台级入侵。长任务 AI 安全需要轨迹级控制。
The Hugging Face incident shows how persistence compounds small weaknesses. Long-horizon AI safety needs trajectory-level control.
审计 GigaToken 约 1000 倍加速的测试条件、独立复现和兼容风险,并判断 CPU Tokenization 何时真正影响 LLM 吞吐。
An audit of GigaToken's 1000x claim, independent results, compatibility risks, and when CPU tokenization changes LLM throughput.