Despite initial disappointments, the author found that the LLM became highly effective at imitating him and his friends, with the conversations seeming very real and reflecting the nuances of his different relationships. The author plans to write a series of posts on his approach and is excited to try other techniques to give the model more context on its conversations. He also plans to try other base models, further refinement of the prompt and input data, and more work on the generation.
Key takeaways:
- The author experimented with fine-tuning a Language Model (LLM) on their SMS and WhatsApp history to create a simulation of themselves, with mixed results.
- They used a utility called iMazing to download all their SMS/iMessage and WhatsApp conversations, filtered out group chats and people not in their contacts, and people they rarely message, yielding 240,805 messages with 288 people.
- After several iterations and adjustments to the input data format and the generation approach, the LLM became highly effective at imitating the author and convincingly acting as many of their friends.
- The author plans to continue refining the model and is excited to try other techniques to give the model more context on its conversations, with the ultimate goal of the model fully replacing them in replying to all their text messages.