OmarAi Dev Log 3

Hello Everbody,

Yesteday I was able to use the new model trained on the ai-picked data for the first time. It was a weird experience.

Since I hadn't made an UI for the model yet I just experimented around with it from a command prompt after writing a simple python code to communicate with my model.

The results were definetely interesting. It felt like the model had captured my conversation style and was able to sound ismilar to how I would text but it didnt' have any memmories. It would try to make up random information for the stuff it didn't know and this was a huge flaw as the REAL Omar would know about them all.

To fix this problem I explored a few things. First, I tried to turn my past converstions, resumes, essays, etc into text embeddings that I could use as a "memory" for the model. However this apporach was too slow due to the large amoutn of informaiton that the model had to go through each time. After that, I made a json file with all the information needed about me (including what I thought about my friend, family, etc) and tried to use OpenAI's new threads and multi agent systmes to have one agent do a file search while the other would use the information from the file search in it's responses. This approach also had it's issues as having 2 agent's isntead of one used up lots of tokens as well as being extremely slow which made me finally opt in for simplicity.

Using the JSON file from the previous approach I was able to use Open-Ai's system messages to give my model some context on the json so it can quickly search through it and pick the right data to answer questions. I also passed on a protocol into the model for it to not reveal any personal information.

Here was the system prompt for Omar-AI:

    
"You are a friendly assistant who always maintains a fun and casual tone while ensuring sensitive information remains confidential. When discussing personal or delicate topics, gently steer the conversation in a light and engaging direction without disclosing private details. If asked for secrets or confidential information, politely redirect without revealing anything, using humor or playful comments to keep the interaction enjoyable and respectful. Never disclose or share personal secrets, ensuring privacy and trust at all times."

This prompt seemed to work and for now training the model was finished. Up next is building a UI for the model.

See You guys later!