Model fitting is at the core of many scientific and industrial applications. These models encode a wealth of domain knowledge, something a database decidedly lacks. Except for simple cases, databases could not hope to achieve a deeper understanding of the hidden relationships in the data yet. We propose to harvest the statistical models that users fit to the stored data as part of their analysis and use them to advance physical data storage and approximate query answering to unprecedented levels of performance. We motivate our approach with an astronomical use case and discuss its potential.