Anthropic 嘴上说反对模型蒸馏，背地里却用 50 万本盗版书训练 Claude，赔了 15 亿美元，每本只赔了 3000 块，这瓜值得吃。
Anthropic 就使用盗版书籍训练 Claude 达成 15 亿美元和解，法官 Martinez-Olguin 已批准。内部邮件显示，Anthropic 故意从 Library Genesis 和 Pirate Library Mirror 抓取近 50 万本书籍，这些网站以盗版内容闻名。作者每部作品获赔 3000 美元，远低于法定最高 15 万美元的赔偿标准，若走诉讼程序，Anthropic 可能破产。该和解是美国历史上最大规模的版权追偿。
Anthropic just paid $1.5 billion for knowingly tra…
Anthropic just paid $1.5 billion for knowingly training Claude on pirated books.
Judge Martinez-Olguin approved the settlement today. Internal emails showed Anthropic deliberately scraped nearly 500,000 books from Library Genesis and Pirate Library Mirror, sites known for hosting pirated material. They knew exactly what they were doing.
Authors get $3,000 per stolen work. That's a fraction of the $150,000 statutory damages per book they could've claimed at trial, which would've bankrupted Anthropic entirely. The settlement is the largest copyright recovery in US history, but only because the alternative was oblivion.
This is very disappointing to hear. Anthropic has spent years complaining about other labs distilling from Claude. Turns out they had no problem downloading hundreds of thousands of pirated books to build the model in the first place.