Day 10 of the remote ML / GenAI internship. Day 9 drew a line. A line can say 200, or −40. A class cannot. I watched Big Brains’ logistic regression one-shot. What I kept: The job is binary: yes or no, 0 or 1. Spam or ham. Fraud or not. Buy or don’t. Same first step as the line: z = w·x + b. Then sigmoid: σ(z) = 1 / (1 + e^(−z)) That squash keeps the answer between 0 and 1. A threshold (usually 0.5) turns the probability into a class. MSE was the marking scheme for a number. Cross-entropy is the marking scheme for a class. Gradient descent still moves the weights. Days 5–7 were the rules. Day 9 was the first model. Day 10 is the first classifier. Question I still have: if fraud is 1 row in 100, is 0.5 still a fair line?