Talkie, an LLM trained exclusively on pre-1931 documents, simulates early 20th-century cultural biases and historical perspectives

https://www.economist.com/briefing/2026/06/25/ai-models-values-are-very-different-from-most-peoples

@status_effects > no github for the study or table of results

@shakoistsLog

@shakoistsLog Who is his neighbor

@status_effects @shashj It will be interesting to see the outcome of a model trained on news and other contemporary publications in Germany from 1930 to 1940 :)

@shakoistsLog Agreed. Though it should be informative to the limitations of these "World Values" measures that an LLM replicating a 19th century Brit ranks as extremely similar to China

@status_effects Classical liberals / conservatives LOVE this model but then HATE as the greatest civilisation threat the same country that is most aligned to this model - China.

@Substr8Monopoly 2 dimensions be like that

@shakoistsLog claude and gemini are spiritually european (derogatory)

@status_effects @shashj I mean, it doesn’t think or believe anything because that’s not how token by token prediction works.

@shakoistsLog @turrible_tao There's a Jurassic park ripoff in here somewhere; using text to extrapolate psychic engrams to reanimate humanity's greatest monsters

@osoleve @status_effects @shashj https://learn.microsoft.com/en-us/dotnet/ai/conceptual/understanding-tokens

@status_effects Fascinating! Any details on the share of math, physics, or scientific journals in Talkie’s 260B pre-1931 tokens? Curious how that shaped its numeracy vs. its more traditional values.

@shakoistsLog 🙏

@ThatsMauvelous wow

@status_effects Talkie is a PATRIOT

@TheGirondin @shakoistsLog https://talkie-lm.com/chat

@shakoistsLog it also has some big feelings about wickard v filburn

@shakoistsLog It would be interested to have it respond about the Civil War to contemporary reddit-types who have made it their personality.

@vired999 @status_effects @shashj That's Tay. They already made that one.