I've been working on a project called Chumi that uses AI to simulate conversations with historical figures — over 3,500 of them, from Einstein and Cleopatra to Confucius and Ada Lovelace.
The core challenge was making each figure feel authentic rather than generic. A few things I learned along the way:
Context grounding matters more than model size. Feeding biographical context, era-specific knowledge, and personality traits into the prompt produces far more convincing results than just asking a large model to "pretend to be Napoleon."
Historical accuracy vs. engaging conversation is a real tension. Users want to ask anachronistic questions ("What would Einstein think about ChatGPT?") but you also don't want the AI making up fake historical claims. We settled on letting figures speculate clearly while staying grounded in what they actually knew/believed.
Multilingual support was harder than expected. We support English, Chinese, Spanish and more — but historical figures spoke in very different registers and styles. Translating personality across languages is a non-trivial problem.
Education turned out to be the killer use case. We built this as a general interest tool, but students and teachers became our most engaged users. There's something about "talking to" a historical figure that makes learning stick.
The platform is free and requires no sign-up: https://www.chumi.io
Curious what this community thinks about the approach. Has anyone else worked on persona-grounded AI systems? What techniques have you found effective for maintaining character consistency across long conversations?
[link] [comments]