Abstract
Proteins are the dominant functional molecules on earth, and yet our ability to leverage them to perform new functions that would be useful to people has largely relied on copying and paraphrasing nature. What does it take to build learning systems that can generalize to new parts of protein space? Amidst the flurry of activity in applying generative modeling to protein design in recent years, I will share some of our own experiences with building learning systems that can generalize, scale, and be programmed to build fit-for-purpose protein complexes on demand.