LLMs encode context, which is only a very weak proxy for meaning. This is why they hallucinate so easily.
Replies
In the most literally sense possible, they absolutely encode the semantic meaning of the language they are given. It's how they work. datasciencedojo.com/blog/embeddi...
apr 27, 2025, 1:12 pm • 0 0 • view
Embeddings model contextual similarity, which is also a flawed proxy for meaning. How good it is depends on how comprehensive your training data is in containing the meaning you're trying to capture
apr 27, 2025, 1:49 pm • 0 0 • view