CrowdStrike's blog discusses its exploration of using overfit machine learning models to detect malicious activity, challenging the traditional emphasis on avoiding overfitting to ensure model generalization. In cybersecurity, the complexity and long-tailed nature of data make it difficult to know if large datasets are sufficient, prompting CrowdStrike to experiment with boosted tree models that memorize training data. Their findings reveal a phenomenon called "double dip," where model performance initially degrades with overfitting but then improves, suggesting that overfit models may outperform traditional models in some contexts. Although preliminary results show promise, the overfit models did not yet surpass benchmark models with regularization and early stopping, highlighting the need for further research and experimentation to optimize hyper-parameters and investigate the potential of interpolated models in cybersecurity applications.