Understanding Vintage Large Language Models: AI Trained on Historical Data

Summary

This article presents a transcript of a talk about "vintage large language models" - LLMs trained on historical data from specific time periods like Roman times, Tudor era, 1960s, or recent decades. The talk introduces the concept of creating language models that reflect the language, knowledge, and perspectives of different historical periods, essentially sending AI back in time through training data selection. The content explores how these vintage LLMs could provide insights into historical language use, cultural contexts, and evolving knowledge systems.