BONUS EP: What If AI Only Knew the Past? Meet Talkie, the Language Model Born in 1930

In this bonus episode we’re going deep on one of the most fascinating AI experiments we’ve come across — a 13 billion parameter language model called Talkie, trained exclusively on 260 billion tokens of text published before 1931. No internet, no modern science, no World War II. Built by former OpenAI and Anthropic researchers, Talkie can hold a full conversation, write working Python code despite never encountering a computer, and reason about history from the inside out. We break down the concept of Vintage LLMs — what they are, why researchers are building them, and what they reveal about AI that modern models simply can’t. From backtesting AI forecasting on real historical events, to asking a model trained before Einstein’s fame to independently derive General Relativity, the implications go way beyond a novelty experiment. This one will change how you think about what AI actually knows — and what it doesn’t.​​​​​​​​​​​​​​​​