Jack, author of a new paper on arXiv, discusses a method for translating text embeddings between vector spaces without paired data, a feat some experts believed impossible. This has implications for embedding security, as it demonstrates that embeddings are not inherently encrypted, even without model access. The method builds upon the "Platonic Representation Hypothesis" and applies to models trained on large, diverse internet datasets. The research tackles a harder problem than simply aligning embeddings of the same data; it aims to generate a comparable embedding in a different model's space given an embedding of unknown text. Questions arose regarding detecting concepts encodable in one model but not another, and potentially modifying models to prevent representation of specific ideas. While alignment methods exist, the authors argue their approach is unique in translating embeddings without requiring seed dictionaries or pre-matched data. Concerns were raised about the generality of the name "vec2vec" and the lack of comparison to existing alignment algorithms. The author also acknowledges the possibility of watermarking embeddings but acknowledges the tradeoff with embedding quality.