What can we learn about from analyzing eye-tracking data from thousands of infants? I will present results from Peekbank (https://peekbank.stanford.edu/), a team science project which aggregates infant eye-tracking data on infant word recognition across many studies to investigate development of word recognition. The talk will focus on three main themes from ongoing work: (1) Aggregating existing data can be powerful for modeling development: by combining data from many experiments, we can overcome the limitations of small, isolated studies to model gradual item-independent changes in online word processing ability across development. (2) Large-scale databases can help researchers make more informed design and modeling decisions: using the Peekbank database, we can systematically explore the impact of a variety of design and modeling decisions, including the effect of selecting longer vs. shorter analysis time windows on establishing reliability. (3) Big data is sometimes not enough: even in the large Peekbank database, word-specific developmental trajectories remain difficult to capture due to high variability within items and idiosyncrasies across individual datasets. Together, these results will highlight opportunities and limitations of current big data approaches to word recognition and infant development, while pointing to future directions that could broaden the usefulness of large-scale databases of infant cognition.