Over the past few years, the Python community has largely unified around the new backtracking pip resolver, but many widely-used ML frameworks which bundle large amounts of binary code have historically pushed at the boundaries of pip's performance envelope and continue to require further innovation. Starting in 2019, I began to investigate how to reduce the size and bandwidth requirements of ML models deployed by the Twitter Cortex ML team, which produced initial drafts of the work that would later be upstreamed into pip as the install --report and --use-feature=fast-deps features. In this talk, I walk through the motivating use cases from Twitter, how these ideas were over time collectively translated into coherent standards, and how to take advantage of these improvements when building Python applications.