#ml

What did I do this Sunday? I went public with my first homegrown model — and added a demo Space where anyone can try it out for free, thanks to Hugging Face.

The initial plan was simple: run a test over all 60,000 colors and see how good the new model really is. The plan didn't hold for long. By the end of the day two models were competing — and even the original dataset was on trial.

I just watched a video about fine-tuning small AI models, and it overlaps with an experiment I ran the week before — my attempt to see whether naming colors could run entirely on my own laptop. After watching the video, I had one question: how many parameters would it take to do the job?

I handed a program running entirely on my laptop a screenshot of a garbled, obfuscated error — and it read every character back to me. Then I asked it to sketch the screen in ASCII art, and it did that too.