Social media data, particularly from platforms like Twitter, have become the largest record of emotions, cognitions, and behaviors in human history, providing a cost-effective and potentially real-time means to assess well-being on a large scale. The methods for assessing well-being from social media data have improved over the last decade to the point where the state-of-the-art now achieves county-specific measurements by week that show equivalent and in some respects superior validity than traditional survey techniques. These advances have occurred largely from two lines of improvement: more robust data aggregation and more accurate psychological estimates drawn from unstructured language.  When viewed as generations of improvement,  this talk will propose that we are now entering a third generation of techniques that is able to leverage within-person longitudinal designs to enable quasi-experimental research designs and achieve unparalleled assessment accuracy.