Researchers built Talkie, a 13-billion-parameter AI trained only on pre-1931 public-domain texts. It uses 234 billion pages of English material frozen at that date. The team tested its reactions to 5,000 events that happened after 1930. Talkie showed surprise at the internet, smartphones, television and the space race. It failed to predict major wars or the rise of the Nazis. It produced plausible Python code despite having no concept of computers. It also hallucinated facts about the past. The project explores how AI models form self-images and interpret the world. A chatbot version is available for public testing. The work also examines how vintage AI can clarify old legal language.

• 13 billion parameters
• 234 billion pages of training data
• 5,000 post-1930 events tested