In the "Attention Is All You Need" paper (the mechanism used to power LLMs) the authors mentioned that they first started working on it for machine translation