It's real embeddings (MiniLM, truncated to 128-dim), so the cosine similarities between tweets are genuine, but it's obviously not X's actual trained model. The

It's real embeddings (MiniLM, truncated to 128-dim), so the cosine similarities between tweets are genuine, but it's obviously not X's actual trained model. The relative distances are meaningful, though.