kogasa, 1 year ago

Semantic embeddings are a thing. LLMs “work with tokens” but they associate them with semantic models internally. You can externalize it via semantic embeddings so that the same semantic models can be shared between LLMs.