Software has been "eating the world" for the last ten years. In the last few years, a new phenomenon has started to emerge: machine learning is eating software. That is, machine learning is radically changing how one builds, deploys, and maintains software — leading some to use the loosely defined phrase Software 2.0. Rather than conventional programming, Software 2.0 systems often accept high-level domain knowledge or are programmed by simply feeding them copious amounts of data.
In this Stanford HAI seminar, Stanford associate professor of computer science Chris Re describes the foundational challenges that these systems present including a theory of weak supervision, guiding self-supervised systems, and high-level abstractions to monitor these systems over time. This builds on his experience with systems including Snorkel, Overton, and Bootleg, which are in use in flagship products at Google, Apple, and many more.
0:43 Speaker Introduction
2:05 Talk overview
2:45 Software 2.0: how ML is changing software
8:50 SW 2.0 product system examples
9:40 Key claim: AI engineering as a discipline
11:47 Overton example: declarative approach to modeling
25:19 Bootleg example: the “tail” problem
30:06 Medical Imaging example: new challenges from high-level abstractions & early progress
41:07 Q & A