AI researchers release “talkie,” a language model trained on 1930s text

A newly released language model called “talkie” offers a striking glimpse into what artificial intelligence looks like when trained exclusively on historical text. The model, described as a 13-billion-parameter system built on data from around 1930, produces responses that blend period-accurate knowledge with confident factual errors.

The official announcement explains that the project presents talkie as a kind of vintage AI, one whose outputs reflect the assumptions, style, and knowledge of a century-old world.

When asked to summarize the history of cinema, talkie produces a detailed but heavily fabricated account. It traces film’s origins to 1550 and a figure called “Baptista Porta,” mixing genuine historical names such as Louis Daguerre and the Lumière brothers with invented events and incorrect dates. The Lumières’ invention of practical cinema, widely dated to 1895, is placed at 1885 in talkie’s account. Researchers note this pattern as characteristic of how the model handles topics it only partially knows.

The model performs far more reliably on topics central to the social world of its training data. Asked about etiquette for calling upon a lady at home, talkie delivers precise, period-accurate advice. It specifies that a gentleman should call between two and six in the afternoon on weekdays, leave his own card regardless of whether he is received, and limit visits to once a fortnight in winter. The tone and detail closely match genuine Victorian and Edwardian etiquette guides.

Talkie also demonstrates a dry, period-appropriate sense of humor. Asked for a dinner party joke suitable for 1920s England, it produces a parliamentary quip about the Houses of Commons and Lords admiring each other for their differences. A second anecdote features a country squire and a clergyman, written in the voice of a Punch magazine column.

The researchers behind talkie observe a clear pattern in its outputs. On well-represented genres such as etiquette guides, humor writing, and formal prose, the model produces convincing and stylistically consistent text. On factual history, it fabricates plausibly but inaccurately, filling gaps with invented names and dates that fit the narrative shape it expects.

The project raises broader questions about how training data shapes an AI’s worldview. Talkie does not know about events after 1930, refers to current cinema attendance figures that reflect the early film era, and applies social rules that have long since disappeared. In this sense, the model functions less like a modern AI assistant and more like an artifact.

Talkie is a 13-billion-parameter model, a size that places it in the same general category as many widely used open-source language models today, though its vintage training data makes direct comparisons difficult.

The project is presented as both a technical experiment and a cultural one, testing what an AI absorbs when its entire frame of reference is a world that no longer exists.