The post discusses recent advancements in aligning vector spaces without needing paired data, as achieved by the authors over two years. They highlight the misconception some experts had about the necessity of paired data for this task, emphasizing that current embedding models, while learning similar outputs, raise significant security concerns. Notably, it has been demonstrated that embeddings can potentially be reversed, leading to risks of information leakage and unauthorized decoding. The discussion prompts questions about the security practices within the industry, particularly in relation to embedding security and methods to resist translation and inversion attacks, which could affect companies dealing in legal or sensitive data.