Day 12 of the remote ML / GenAI internship. Parameters I already knew. Accuracy I thought I knew. I watched two Big Brains videos — model parameters, then accuracy and the confusion matrix. Parameters are learned: weights, biases, coefficients. Hyperparameters are set by me: learning rate, depth, k. Do not tune those on the test set. Accuracy is exam percentage: correct predictions / total predictions. 80 out of 100 looks fine. Until the table is uneven. Video case: 800 dogs. 200 cats. Model says “dog” every time. Accuracy: 80%. It never found a cat. The confusion matrix opens the score: TP — said yes, was yes TN — said no, was no FP — said yes, was no FN — said no, was yes Accuracy = (TP + TN) / all. It hides which cell you failed. Same rule as Day 5: a single number is not proof. Question I still have: if fraud is 1 in 100, which cell should hurt more — FP or FN?