Google just dropped Gemini Embedding 2, and I’m genuinely excited about this one because it’s a much more sensible tool for a bunch of problems people keep throwing LLMs at. In this video, I break down the difference between a large language model and an embedding model in plain terms: an LLM is a next-token predictor, while an embedding model is a representation model. You feed it text (and now more than text), it spits out a vector, and that vector lets you measure similarity between items using simple distance/similarity math. I walk through how embeddings power semantic search: you precompute vectors for your dataset, store them, and then at query time you embed the user’s input and run something like cosine similarity to rank results. That’s faster and way cheaper than asking an LLM to compare every single item on every request. The big news here is multimodality—Gemini Embedding 2 can embed text, images, video, audio, and documents, and it can even take interleaved inputs (like image + text) in one request. Google also mentions MRL (Matryoshka Representation Learning), which lets you scale output dimensions down from the default 3072 so you can trade off quality vs storage and cost.