Why You Can't Trust Your Training Loss

Four consecutive checkpoint saves, 250 steps apart: 1.11 → 0.58 → 1.32 → 0.49 A 63% swing between adjacent saves. The trend is real. The step-to-step reading is noise — and the trend went down for six and a half days while the model got steadily worse. I trained an AI on my own face for a week to find out exactly where it broke. Full video: https://youtu.be/w_7EpghF8fw #MachineLearning #AITraining #LoRA #DeepLearning