Serverless computing promised a boundless programmable cloud, but delivered an army of incommunicado amnesiacs. Still, serverless computing is a hint of a better future: clouds as globe-spanning, auto-expanding supercomputers for anyone to program. In this talk I'll share lessons from research—both foundations like the CALM Theorem and practical experience from open source like the Anna KVS prototype—on how we can do so much more to deliver stateful, communicating, autoscaling cloud software. Then I'll describe our ongoing work in the Hydro Project at Berkeley, which looks ahead to a fully-featured programmable cloud, powered by a new data-centric compiler stack.