When traditional study materials vanished in grad school, I turned to AI to generate synthetic practice exams. My undergraduate experience at a US university had conditioned me to expect abundant past exams, especially for foundational courses like Calculus and Linear Algebra. I'd solved dozens of decade-old problems to prepare, treating them as essential training data.

Moving to Germany for my master's degree brought an unexpected challenge: past exams were rarely available. With finals approaching and no practice problems to solve, I faced what I called the 'Human Training Data Problem.' Like a machine learning model starved of training data, I couldn't bridge the gap between lecture notes and potential exam questions. The situation felt dire, especially with back-to-back finals looming.

I applied synthetic data principles from AI research to create my own practice materials. Using Claude, I generated mock exams based on exam structure hints from professors and upperclassmen. For formulaic subjects, I replicated question styles with slight variations. The approach worked - I avoided going in blind and maintained the problem-solving practice that had served me well during my undergraduate years. Sometimes the most effective solutions come from treating human learning like machine learning.