A college student experimenting with AI language models got more than he bargained for when his tiny Victorian-themed AI surprisingly mentioned real events from 1834 London. The developer, Reddit user Hayk Grigorian, built a small model called TimeCapsuleLLM that was trained solely on texts from London between 1800 and 1875. His goal was to capture the language style of the Victorian era, full of biblical references and period-appropriate rhetoric. But one simple prompt led the AI to produce a detailed account of protests from 1834, catching Grigorian off guard.

How a Small AI Discovered Real 1834 Events

Grigorian’s AI was designed to continue whatever text he fed into it. When he prompted it with “It was the year of our Lord 1834,” the model generated a paragraph about protests filling London streets. The text mentioned Lord Palmerston, who was involved in the events of that year. Curious, Grigorian looked up Palmerston and found that he was indeed a key figure during the protests sparked by the Poor Law Amendment Act of 1834. The AI’s output was surprisingly accurate and connected historical dots that Grigorian hadn’t explicitly programmed into the model.

Training the AI with Victorian Texts Only

Instead of fine-tuning existing modern models, Grigorian trained his AI from scratch using only Victorian-era sources. He collected over 7,000 books, newspapers, and legal documents from London published in the 1800s. To keep the language authentic, he created a special tokenizer that removes modern vocabulary, ensuring the AI only learns Victorian language patterns. This process, which he calls “Selective Temporal Training,” resulted in a model that speaks like someone from that time period.

From Gibberish to Genuine Historical References

In earlier versions, the AI produced mostly nonsense or made-up stories that sounded Victorian but weren’t real. As he scaled up the data and model size, the outputs became more coherent. The current version, with 700 million parameters, now occasionally references actual historical figures and events, like Lord Palmerston and the 1834 protests. The improvements are typical in AI development—larger, better-trained models tend to “remember” more from their training data and produce more accurate outputs.

Why This Matters for History and Language Study

While these models aren’t perfect and can still invent facts, they open interesting doors for research. Historians and digital humanists could use such models to simulate conversations with people from past eras, helping to understand old speech patterns and vocabulary. Even if the AI occasionally “hallucinates,” these errors can still be revealing about the linguistic style of the time. Grigorian hopes to expand his project to include other cities and languages, inviting collaboration to refine these historical AI models further.

In a world where AI often fabricates details, this accidental “factcident” is a breath of fresh air—an AI that, by chance, is telling the truth about history. It shows how even small, hobbyist projects can stumble upon surprising insights, especially when trained on vast amounts of old texts. As AI continues to evolve, these kinds of experiments remind us of its potential to connect us with the past in new and unexpected ways.

Inspired by

• https://arstechnica.com/information-technology/2025/08/ai-built-from-1800s-texts-surprises-creator-by-mentioning-real-1834-london-protests/

Sources

• conde.digital
• cnevids.com
• github.com
• github.com
• pnas.org