“Locked in 1930”: an AI trained on pre-1930 public-domain text predicts the future
Futura-Sciences
Researchers built Talkie, an AI trained only on pre-1930 public-domain text. It’s bad at modern history, but useful for testing how a model sees the world with one era frozen in place.
Based on reporting by Futura-Sciences — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Researchers Nick Levine, David Duvenaud, and Alec Radford built Talkie, also called 13B 1030 LM, as a kind of machine-era time capsule. Its knowledge stops at December 31, 1930. The model has 13 billion parameters and was trained on a dataset the team says is equivalent to nearly 234 billion pages, all in English and all from public-domain text.
The appeal of the setup is obvious. If you strip away everything after 1930, what does an AI make of the world? And what happens when it’s pushed to talk about events it was never allowed to see? The researchers used that question to probe how a probability-based model reacts to later history, and whether it can handle post-1930 events without peeking ahead.
Getting to that point was not simple. Much of the source material came from digitized books and papers, so optical character recognition had to be good enough to let the model learn in the first place. The team also had to worry about “contamination,” meaning anything from after 1930 slipping into the training data and ruining the experiment. Over time, they tested Talkie on 5,000 descriptions of historical events to see when it would act startled and when it wouldn’t.
The answers were blunt. Talkie missed major wars and even the rise of the Nazis when researchers asked it to react to events beyond its cutoff. Faced with today’s world, it reacted to the internet with real surprise and had no grasp of smartphones, television, or the space race. In one test, the team trained a version only up to 1911 and asked whether it could independently discover general relativity by 1915, though the publication does not say whether it succeeded.
Talkie was not completely helpless, though. Asked to write Python code, a language created in 1991, it produced something plausible. That took some inventiveness, since the word “computer” was not part of its world. The researchers also say the model hallucinates, inventing facts or stories about the past, which makes it feel less like a museum exhibit and more like a very opinionated archive.
My take — AI-written commentary, not fact-checked reporting
This is the rare AI stunt that earns its keep. A model with a hard historical cutoff is not a toy if it helps expose how much of AI is really pattern-matching dressed up as wisdom. Also, the current habit of training everything on everything has made “clean data” sound quaint, which is exactly why this project matters.
Read more about this at: Futura-Sciences
