Shoutout to the tiny AI model that guessed my bakery order wrong, and taught me something about fine-tuning
I run a small bakery in Portland and messed around with a tiny open source model on my laptop to predict daily croissant demand. Fed it two years of sales data and it kept overshooting Fridays by 40 percent, so I added a note about rainy weather and it flipped to undershooting by 30. Turns out the model was just copying whatever pattern I fed it last, which is basically what happens when you fine-tune on too little data. Anyone else run into this where the model seems smart until you give it one weird input and it falls apart?