The concept of sample efficiency in speech recognition is gaining importance as technology advances rapidly. The current paradigm of supervised learning requires large amounts of labeled data, which can be time-consuming and expensive to obtain. This leads to issues such as slow training times, high costs, and potential errors in labeling. In contrast, self-supervised learning approaches aim to reduce the dependence on labeled data by predicting properties of the original data without requiring labels. These methods have shown promise in addressing the shortcomings of supervised learning and moving closer to the lower bound of potential sample efficiency. The implications for industry are significant, with reduced need for data labeling, changes in data cleaning processes, and new business opportunities arising from the availability of unlabeled data. As machine learning algorithms continue to evolve, businesses should be aware that deep learning may not dominate in the future, and instead, a different paradigm may emerge.