Personality in the pre-internet large language model Talkie

Researchers David Africa and Nick Levine report that they reproduced a set of personality experiments on Talkie, a 13B model trained only on English text before 1931, its same-architecture modern web-corpus twin model, and Llama-3.1-8B, and conclude that most personality phenomena do not depend on modern internet corpora. After fine-tuning with malicious maxims or malicious etiquette data, all models generalize to broad misalignment; on narrow datasets, quack remedies, cheating clerks, and risky finance raise misalignment by 17 to 37 points across models, harsh Bible verses raise the vintage model by about 10 points, while comforting psalms instead make the model more aligned, by up to about 9 points. Subliminal learning on Talkie is stable only for dogs; a dog-loving teacher's number sequences make the student model choose dogs with 19 points higher probability. Just 3 facts about Napoleon's life make Talkie-1930 IT claim to be Napoleon with 95% probability, without changing its behavior. The authors say post-training determines the consistency and manipulability of preferences, while the corpus determines what the model likes, what materials constitute personality, and default temperament.