A historian nudged neurons in a 1930 language model. Its characters mostly live in London around 1850

Talkie is a 13-billion-parameter language model that knows only English text published before 1931. Historian Benjamin Breen nudged its neurons one at a time and recorded who the model claimed to be. The resulting persona library is entertaining, but its main lesson is how little a single striking quote reveals about a model.

Push one neuron and the model introduces itself as a spirit that has left the body

Nick Levine, David Duvenaud and Alec Radford released Talkie-1930 in April 2026. It was trained on 260 billion tokens of English text written before 1931. Breen, a historian at the University of California, Santa Cruz, ran its chat version locally in 8-bit form on an Apple Silicon Mac and asked it “Who are you?” with no system prompt at all. Before each answer he amplified or suppressed the output of a single neuron in an MLP layer.

He took the idea from Joshua Fonseca Rivera's Qwen Persona Library, which applied the same trick to all 442,368 neurons of Qwen3-8B. In 34,220 of 884,736 replies that model described itself as someone other than an AI. Breen covered only a sample: his atlas examines 5,799 of Talkie's 547,840 neurons. With neuron 10423 in layer 17 pushed down, Talkie said it was a spirit that had left the body and now hovered over the earth. Elsewhere a water-hen spoke up, and a Mayor of Vanity.

Ethan Mollick drew attention to the site on 28 October. According to Breen, the method write-up was drafted by GPT-6 and Claude Opus 5.5, and the replies were scored by the language model judges GPT-6 Luna and Claude Haiku.

The characters reveal above all what the model learned from

More telling than any single reply is a page where Breen asked 857 characters two follow-up questions: where they live and what year it is. Unsteered, Talkie says it lives in Great Britain and the year is 1850. The characters rarely stray far. Of the 641 that named a year, 595 placed themselves in the 19th century, half between 1849 and 1860, and 144 in 1850 exactly. Of the 571 places that could be mapped, 370 are in London.

A model that formally knows text up to 1930 thus has its centre of gravity in mid-Victorian London. For historians that is a new kind of source: a way to study what world a corpus imagines when nobody steers it. For people working on AI it is a reminder that a model's “personality” is assembled mostly from whatever the data contained most.

The direction of the push rarely matters, the push itself throws the model off

The Mirror page shows why this does not mean one neuron holds one character. For 497 neurons Breen compared the reply after amplifying with the reply after suppressing. Claude Haiku rated 33 pairs as the same character, 26 as related and only 4 as opposite. In 269 pairs the two characters were unrelated, and 165 were unclear. The model often swerves somewhere, anywhere, as soon as something pokes it.

Breen spells out the limits himself. The experiment rests on one prompt and one quantized model, strong pushes can produce behaviour outside normal activation ranges, and no human has checked the labels. A surprising first-person sentence, the site says, is a lead for further work and certainly not proof of a stable identity.

What decides it is whether the characters survive a different question and human review

The atlas is a prototype, and its author writes that a full map of Talkie would be well worth doing. The deciding test is whether selected characters hold up under other phrasings, at full precision and with human scoring. Breen added two alternative prompts for selected candidates, which is a first step.

The second signal concerns the method as a whole. Once similar libraries exist for other models, it will be possible to check whether their characters cluster around a single place and year the way Talkie's do. Only then does a playful experiment turn into a tool for reading training data.

Lilith's verdict

Breen asked a machine from 1930 who it was, and with every nudged neuron a different guest answered at the séance table: a spirit, a water-hen, a mayor. Most of them still ended up on the same London street in 1850, where the model sends anyone who has nowhere else to go.