LLMs are universal translators (2023) (lil.law.harvard.edu)

Recent insights highlight large language models (LLMs) as "universal translators," capable of translating any input to any output by embedding concepts into a single, high-dimensional space. This notion emerged during presentations at a Spanish-language library conference in Argentina, where practical applications of LLMs were showcased. By translating presentation slides in real time and utilizing NLP tools to facilitate communication, LLMs proved to be versatile and powerful translation tools, emphasizing their utility across diverse tasks—from programming to legal brief generation. The significance for the AI/ML community lies in the realization that LLMs can democratize and personalize translation processes, allowing individuals to create tailored solutions that were once the domain of dedicated applications. With billions of tokens feeding into their architecture, LLMs can understand complex relationships in language and other formats, creating new opportunities for innovation. As the landscape of technology evolves, this flexibility of LLMs shifts the paradigm from specialized software to adaptable, individual-centered tools, potentially transforming various industries and enriching human connections through improved communication.