Predicting atmospheric blocking is a challenge in weather and climate modeling due to its complex dynamics and significant impact on extreme events like heatwaves and extreme precipitations. While recent deep learning-based forecasting systems show promise, it’s an open question about their ability to predict unseen extreme weather events, particularly without training on them. This study takes a minimalist approach by using a two-layer quasi-geostrophic (QG) model to simulate atmospheric blocking. We create a "miniature" version of a state-of-the-art machine learning workflow, inspired by the FourCastNet v1 framework, to analyze how ML architectures forecast blocking events. Our findings indicate that in this simplified framework, predictive capability for blocking comes from understanding general flow dynamics rather than direct exposure to blocking examples. This suggests that models do not need specific training on blocking events to make effective predictions. Our results underscore the significance of rare-event representation in ML-based geophysical predictions and highlight important considerations for dataset design, especially in areas where training resources are limited. Physically, our findings suggest that blocking events share similar dynamical processes with many other non-blocking events, suggesting that we do not need to train models specifically on blocking events to predict them effectively.