This Chatbot Is Frozen in 1930
Talkie is a chatbot trained on pre-1930s data that lets you speak with someone who has never even heard of a computer. Hot diggity dog!
What would it be like to travel 100 years into the past and speak with people who had never heard of a computer, the internet, space travel, or even television? It’s hard to say exactly since no one has ever done such a thing. And considering the implications of the Fermi Paradox of time travel, it is unlikely that anyone ever will.
However, thanks to the clever thinking of some AI researchers, we can now all have this experience (sort of). A team led by Nick Levine has created what they call Talkie. It is a 13-billion-parameter vintage language model that was trained from scratch entirely on text from the year 1930 and before. The dataset it was trained on consisted of 260 billion tokens, producing a model that is — as Talkie might say — the bee's knees.
By being deliberately cut off from anything that happened after 1930, Talkie reflects the worldview, knowledge, and linguistic quirks of the early 20th century. Ask it about global politics, and it may speak confidently about interwar tensions — but remain unaware of World War II. Ask it about technology, and it might marvel at radio while having no concept of smartphones.
The developers say this “vintage” approach isn’t just a novelty. It provides a unique way to study how language models understand time, knowledge, and prediction. In one experiment, the team tested how “surprising” historical events appeared to Talkie. Events that occurred after 1930 were, unsurprisingly, harder for the model to anticipate, with a noticeable spike in unpredictability around the mid-20th century.
The project also sheds light on a deeper question: can an AI predict the future based only on past knowledge? While Talkie doesn’t yet offer reliable foresight, early experiments suggest that larger models may gradually improve at extrapolating trends beyond their training data.
Another surprising capability is coding. Despite never encountering computers during training, Talkie can learn simple programming tasks when given examples. In one case, it successfully inferred how to reverse a basic encoding function in Python — hinting at an ability to generalize abstract patterns, even in unfamiliar domains.
Still, the system has clear limitations. Because its training data comes from scanned historical documents, errors introduced by optical character recognition can degrade performance. There’s also the persistent challenge of “temporal leakage,” where stray modern references slip into the dataset, subtly contaminating the model’s historical purity.
If you’d like to chat with Talkie, you can download the model from Hugging Face.