A review day in the remote ML / GenAI internship. Three videos. One rule. A training score is not proof. What I kept: 1. Selection Hold rows back. Train learns. Validation tunes. Test is one honest score. Do not tune on the test set. k-fold is the same idea, rotated. 2. Fit Underfit: bad on train and test. Too simple. Overfit: great on train, poor on test. It memorized the homework. 3. Loss Loss is the miss: actual minus predicted. Square it so it cannot go negative. Average it. A number needs MSE. A class needs cross-entropy. Parameters move to shrink that miss. The three lock: Loss tells the model how wrong it was. Validation tells me whether that move generalized. Overfit is a tiny train miss and a large test miss. Question I still have: if two models have similar test loss, do I pick the simpler one?