AI models have traditionally been trained using internet data, which heavily influences their knowledge base and communication style. However, a research team has ventured into uncharted territory by creating an AI model that deviates from this norm.
Talkie-1930is an open-weight model with 13 billion parameters, developed exclusively from texts published before January 1, 1931, including books, newspapers, scientific journals, patent filings, and legal cases. This selection is not due to budget or data constraints but because these works are public domain in the US.
This AI, devoid of knowledge about the internet, civil rights movements, or the Cold War, offers insights into medicine prior to penicillin’s widespread use and lacks any understanding of computers, cryptocurrency, artificial intelligence, memes, or internet culture. It is currently live at attalkie-lm.com/chat, where Claude Sonnet engages it continuously for public interaction.
The project, led by Nick Levine, David Duvenaud, and Alec Radford with support from Anthropic, aims to combat benchmark contamination—a problem in AI testing—by eliminating any chance of modern data influencing scores. They released two checkpoints under Apache 2.0: a base model for training purposes and an instruction-tuned conversation version.
The team explored how the model reacts to historical events post-1931, noting peak surprises around the mid-20th century. Their research questions how an LLM’s identity is shaped when trained away from internet data. They plan to achieve a GPT-3-level model by summer 2026, with a corpus that could reach over a trillion tokens.
When queried about Adolf Hitler, the AI—limited by its pre-1930 knowledge—predicted his rise to power and speculated on Germany’s potential shift towards monarchy without foreseeing future atrocities. This mirrors early 1930s Western press views, reflecting its training data accurately.
On the concept of “thinking machines,” it envisioned mechanical brains fostering global communication while noting language barriers as a primary challenge. It suggested establishing a universal language to overcome this obstacle but warned against over-reliance on such technology, fearing it might impede natural development and self-sufficiency.
In financial advice, Talkie recommended investing in entities like Canadian Pacific Railway and De Beers—sensible choices for the era’s economic landscape. Despite some companies’ subsequent changes or liquidations, its logic aligned with 1930 investment principles.
Asked to predict 2026, it envisioned a world without standing armies or widespread crime due to education’s proliferation—a reflection of then-visible trends rather than present realities. This prediction underscores how unforeseen historical events have dramatically altered societal trajectories since the 1930s.
Both model checkpoints are available on Hugging Face under Apache 2.0, with local operation requiring a CUDA GPU with at least 28GB of VRAM.