Rajiv Shah on Instagram: "This paper introduces vec2vec, a method that aligns text embeddings from different language models—without access to the models or labeled data. It supports the Platonic Representation Hypothesis, showing that large models trained on different data still learn embeddings that can be transformed into one another. The results have serious implications for vector database privacy, as attackers can reconstruct sensitive content from just 10k embeddings. Harnessing the Unive