Wasted 3 months labeling every training image by hand before a coworker asked why I wasn't using a pretrained model
I was building an image sorter for my side project and hand labeled about 4,200 photos over 3 months because I figured clean data beats everything. Last Tuesday a buddy at the coffee shop in Austin watched me scrolling through labels and asked why I wasn't just starting from a pretrained model and fine tuning on maybe 300 shots. I ran it that night and the thing hit 91 percent accuracy with a fraction of the effort, which stung a little. Turns out I was treating it like a data entry job when the tools had moved way past that. Anyone else get stuck doing grunt work on an AI project because you didn't check what the off the shelf stuff could already do?