Published on

A developer replaced his best friends with an LLM trained on 500k group chat messages


A developer by the name of Izzy Miller trained an uncensored large language model on a college-era group chat that they and their best friends still use, utilizing LlaMa, Modal, and Hex. The dataset for this project consists of their Group Chat, which includes conversations with their five closest college friends. Despite living in different parts of the country, the group chat has remained active for the past 7 years, accumulating 500,000 messages. To replicate this process with one's own chat.db, it is possible to clone the Hex project and follow the process mainly automatically. However, be cautious as this requires uploading the chat.db to the cloud. While Hex is a secure platform, some might prefer to perform the process locally. It was more convenient for Izzy to figure out schemas and queries using Hex initially, but running the code locally should be a simple copy/paste task.

Izzy began this project soon after the Stanford Alpaca project released their code for fine-tuning LLaMa, which appeared to be an ideal choice for a small homebrew model. Although it was state-of-the-art three weeks prior, there are now numerous other projects for running affordable small LLaMa-based LLMs, such as llama.cpp and Alpaca-LoRa. It is worth spending a few minutes browsing to see if there is a more suitable model for one's specific needs.

The project has genuinely provided Izzy and their friends with countless hours of deep enjoyment. The training process seemed to optimize for outrageous behavior, and observing their conversations from a third-person perspective highlighted the amusing and ridiculous nature of the exchanges. The model successfully captured the voice and perspectives of Izzy's friends while retaining a substantial amount of information about their preferences and lives. Although Izzy considered attaching an embedding database like Chroma to provide the friends with a knowledge store, they ultimately found this unnecessary. The AI model already knew who each person was dating, their interests, and other personal details. Izzy highly encourages others to clone this project and follow the tutorial or undertake a similarly complex yet seemingly pointless AI project. It serves as an excellent introduction to AI and offers a firsthand experience with the powerful technology that has people discussing potential doomsday scenarios.


Hacker News Discussion: