Safety

Word2Vec and Word Embeddings Word embedding 把离散的符号（词）映射为 dense vectors，使语义关系可以用 vector space 中的距离和方向来表达。本文沿着 CS224N Lecture 2 的思路，依次梳理 Word2Vec、Negative Sampling、count-based methods、GloVe、evaluation、word senses 以及基于窗口的神经网络分类器；最后补充 Transformer 如何把初始的 token embedding 转化为 contextual representation。 1. One-hot Encoding 词表大小为 $V$ 时，第 $i$ 个词的 one-hot vector 为： $$ e_i=[0,\ldots,0,1,0,\ldots,0]\in\mathbb{R}^{V} $$这是一种 localist representation：每个词独占一个维度，vectors 之间互不重叠。它只能表示词在 vocabulary 中的 identity，既不包含词在句子中的位置信息，也不包含词与词之间的语义相似度。 不同词的 one-hot vectors 彼此正交： $$ e_i^\top e_j=0,\quad i\neq j $$因此在 one-hot 空间中，cat–dog 与 cat–car 之间的"距离"完全相同，无法体现两者语义相似度上的差异。 2. Word2Vec: Learning from Local Context Word2Vec 的基本假设是 distributional hypothesis：经常出现在相似上下文中的词，通常具有相似的语义。 ...

What is this A system card is a technical document Anthropic publishes for every Claude release, documenting capabilities, safety evaluations, and deployment decisions. Opus 4.7’s card (April 2026) is 232 pages — unusually thick. This note explains what fills those pages, and what that implies about how Anthropic actually operates. The short answer It’s not a “model card” in the academic-paper sense. It’s a cross-audit: nine orthogonal risk/capability domains, each with its own evaluation methodology, external verifications, failure-case transcripts, and comparisons against both the prior model (Opus 4.6) and a more capable internal model (Mythos Preview). Each domain could be its own paper; they bind them into one document so the full picture is visible. ...