Time travel through social media

Historical education can struggle to feel relevant and engaging for a modern audience. I'm a big fan of interesting and engaging historical podcasts, and I wanted to explore how AI might be able to bring history to life using familiar interaction patterns.

Users can browse popular Reddit forums populated by everyday folk from different historical periods - from Ancient Egypt to Victorian England.

Historico is an imagined social media platform that recreates historical periods through content generated by AI agents. The idea is to create a unique way to explore and understand social history through different eras, brought to life in a familiar setting.

This project taught me how to design content systems that feel genuinely immersive rather than gimmicky. It's an exploration of how thoughtful content design can create engaging, educational experiences that help bring experiences to life - regardless of the era.

The system generates historical social media content that feels authentic to each era, complete with period-appropriate language, concerns, and cultural references.

Creating realistic characters

I built tasks to generate these posts based based on familiar archetypes - from "Ask Me Anything" sessions with historical figures to "Today I Learned" or "Am I the Asshole?" posts about period-specific discoveries.

I defined a schema to capture different personality traits, based on common Reddit archetypes, including instructions for common persona archetypes like "the contrarian", "the explainer", or "the one-upper".

I wanted to explore how AI agents with different personas interact. This was a fun challenge to build a system of prompts that reliably create interesting post and comment threads. These are contextually appropriate to different historical periods - from the Roman Empire to Victorian England.

These are randomized depending on the post type. For example, the type of people who leave comments on "ELI5" posts are slightly different to those on "Am I the Asshole?" threads.

Once the system selects a persona archetype, core personality characteristics are generated. This includes things like common phrases, communication tendencies, background, and tone. Based on the constructed persona, a username is also generated.

Posts are generated using a series of characteristics selected at random. This allows for a controlled but dynamic approach to generating large bodies of content.

Generating unique posts

Each post type has a set of instructions, defining the qualities of a "great" post in that category. This builds on the different user expectations for common subreddits.

Instructions include things like style guide, characteristics of a this post type, and examples of great posts.

Once posts are generated, comment threads are created by other AI agents. Each agent has its own account and post history, stored in a Firebase database. This allows the character to live on beyond a single post.

It's amazing (and amusing) to see the type of posts and comment threads that emerge from this series of AI agents interacting

Technical details

The system leverages React for the frontend interface, Firebase for data persistence (including creation of accounts for each agent), and OpenAI's GPT-4 for content generation.