Ask HN: How to balance theory and practice when learning Machine Learning?

On one hand, delving into academic papers and fully understanding the math involved seems to demand a commitment akin to pursuing a PhD in Machine Learning. I'm also not entirely sure how important/useful this approach is. Especially for those not involved in cutting-edge research. On the other hand, the engineering approach emphasizes empirical experimentation and software engineering practices: fine-tuning, hyperparameters, APIs, MLOps pipelines, etc. From this approach, the models feel like black boxes with some configurable settings. How do you balance these two aspects when learning about Machine Learning? How do you decide which to focus on, especially given how fast the entire field is changing month by month?