College student’s “time travel” AI experiment accidentally outputs real 1834 history
1 min read
Summary
Hayk Grigorian, a self-described ‘hobbyist developer’ is building AI language models that reproduce Victorian-era English, with the ultimate goal of capturing an authentic Victorian voice in the model’s outputs.
To achieve this, Grigorian has been developing a model dubbed TimeCapsuleLLM, which he has trained on texts from 1800 to 1875 London, and which produces text heavy with biblical references and suitable rhetorical styles of the era.
This fits into a broader field of similar ‘Historical Large Language Models’ (HLLMs) that have been developed, such as MonadGPT, which can discuss topics using 17th-century knowledge frameworks, and XunziALLM, which generates classical Chinese poetry.
In testing the model, Grigorian was surprised to learn that the AI’s output referred to real protests that took place in 1834 London, suggesting that the model had accidentally learned details about historical events that the developer himself was previously unaware of.
