HAL9000
I am old enough to remember life before the Internet. Now I say I am old enough to remember life before Artificial Intelligence. That's how profound I think the impact A.I. will have on our lives. And we're still in the "AOL" stage of A.I.!
HAL9000 from the movie 2001: A Space Odyssey is the archetype of A.I.: capable, wired into everything, and it cannot be trusted. So I selected "Hal" for my first "real" A.I. project. In hindsight, I realize that it's probably the obvious choice for a lot of people. But I went into it blind: I did not consult anyone's notes on building their "Hal" as I developed my own.
Voice Model
Training the voice model was the biggest task of the project. I started with the Coqui TTS library. But I had a lot of trouble with dependencies because the project has been abandoned. I switched to F5 TTS, which was much easier. I used the Gradio UI to train the voice on audio clips I created from the movie.
Hugging Face LLM
This project introduced me to Hugging Face. I was surprised when I learned of it. It can't be this easy, I thought. You just seach for the model you want and download it? And it comes with its own knowledge base built in?
In this stage, I learned about finetuning an A.I. with properties like temperature and nucleus sampling. High values were needed to make Hal creepy and imaginative.
User Interface
The user interface was the relaxing part. I think of it as drawing or illustrating, one of my favorite activities. I built a website using Next.JS and Tailwind. I am proud to say that the HAL9000 "illustration" is completely Tailwind classes - no images necessary. I borrowed the code for the "transcript monitor" from the Role Matcher feature on this website. I think the result compares with its movie muse nicely.