AI 导读
Engram 为短 token 序列检索已学到的向量，帮助模型复用熟悉的模式，而不是重新构建它们。 我们的扫描发现了诸如"Ian Goodfellow"这样的名字、代码片段、任务指令和网站样板文本。(2/6)
正文
Engram retrieves learned vectors for short token sequences, helping the model reuse familiar patterns instead of reconstructing them.
Our scan found names like “Ian Goodfellow”, code fragments, task instructions and website boilerplate. (2/6)
来源：@SemiAnalysis_ · x.com