Embeddings
Dense vector representations that capture semantic meaning. The bridge between human language and machine understanding.
Embeddings map discrete symbols (words, tokens, entities) into continuous vector spaces where geometric relationships encode semantic ones. Word2Vec showed that vector arithmetic captures analogies: king - man + woman = queen. Contextual embeddings (ELMo, BERT) go further — the same word gets different vectors depending on context, resolving polysemy. Embeddings are the foundation of modern NLP, recommendation systems, and retrieval-augmented generation. The remarkable fact is that meaning, which seems ineffable, can be approximated by position in a high-dimensional space. Similarity in vector space maps reliably to similarity in meaning.