Anthropic 封禁数万蒸馏 Claude 的账号
Anthropic 称有公司偷偷用 Claude 蒸馏训练自家模型，已封禁数万个相关账号，常在两小时内动手，部分正常开发者连接方式也受影响。
Anthropic just blocked rivals from copying Claude's intelligence 🤯 Companies were secretly asking Claude millions of questions, saving its reasoning, and using that to train their own cheaper models. Without permission, New defenses cut the data copycats can steal by 90 to 99%. GLM used this to build one of the most hacking-capable AI models anyone can download. Anthropic says its safety controls are easy to bypass, Moonshot's Kimi ran a campaign that included a request linked to China's military analyzing surveillance footage of a specific person. Anthropic banned tens of thousands of accounts tied to these operations. Often within 2 hours. The catch. Some real developers may need to update how they connect to Claude. A small number of normal requests still get blocked. Quietly copying a top AI model's thinking to build a cheaper version just got a lot harder
这条是引用着另一条写的：x.com/1353836358901501952/status/210979589890995…