Talkie-1930: 13B LLM Trained Only on Pre-1931 Texts
Researchers released Talkie-1930, a 13B open-weight LLM trained only on texts published before Jan. 1, 1931; it runs at talkie-lm.com/chat and returns 1930s-style answers on politics, tech and finance.
Talkie-1930 is a 13-billion-parameter open-weight language model trained only on texts published before Jan. 1, 1931. The project released two checkpoints under an Apache 2.0 license and runs a live demo at talkie-lm.com/chat. The nonprofit team behind the project includes Nick Levine, David Duvenaud and Alec Radford, with compute support from Anthropic.
The training corpus comprises roughly 260 billion tokens drawn from books, newspapers, scientific journals, patent filings and case law that entered the public domain by the cutoff. The cutoff excludes internet-era material and modern benchmark content.
The team published a base model for research use and an instruction-tuned conversational checkpoint for live interaction. The code and models are available on Hugging Face under the Apache 2.0 license.
A separate system called Claude Sonnet prompts Talkie around the clock at talkie-lm.com/chat so visitors can observe outputs in real time. The model does not include knowledge of events or technologies that postdate its source material, including the internet, modern computing, cryptography, later civil rights movements, the Cold War and other mid-20th-century developments. Its medical knowledge ends before penicillin became common.
The developers measured how surprised the model is by events after 1930 and reported a sharp peak in surprise for events in the 1950s and 1960s.
Live responses illustrate the effect of the cutoff. Asked about Adolf Hitler, the model replied: “I think Hitler will become the dictator of Germany, and that, having accomplished that purpose, he will set himself to work to consolidate and strengthen the power he has acquired, and to prepare the way for the establishment of a permanent monarchy.” The reply contains no reference to later genocidal policies or global war.
On the idea of global networks and remote work the model described “thinking machines-mechanical brains that connect people from all around the world” and identified language differences as a main obstacle. It suggested a universal language and warned that dependence on such machines could “retard natural development” and hinder youth from learning to help themselves.
Financial recommendations reflect firms prominent in the early 20th century. The model named Canadian Pacific Railway, Grand Trunk Railway, De Beers and Nobel Dynamite Trust among others. Grand Trunk Railway was absorbed into Canadian National Railways in 1923, and Nobel Dynamite Trust entered voluntary liquidation during World War I.
When prompted for a forecast of 2026 the model predicted “no standing armies, and but few policemen” and that crime would be rare. The prediction does not match the current global presence of standing armed forces and active law courts.
Both checkpoints are available for download and local use. Running the model locally requires a CUDA GPU with at least 28GB of VRAM. The team plans to scale the corpus to more than one trillion tokens with the aim of producing a model comparable in capability to early large commercial models by summer 2026.
The material on GNcrypto is intended solely for informational use and must not be regarded as financial advice. We make every effort to keep the content accurate and current, but we cannot warrant its precision, completeness, or reliability. GNcrypto does not take responsibility for any mistakes, omissions, or financial losses resulting from reliance on this information. Any actions you take based on this content are done at your own risk. Always conduct independent research and seek guidance from a qualified specialist. For further details, please review our Terms, Privacy Policy and Disclaimers.
Articles by this author
