Machine learning is quickly becoming a dominant force at Facebook, and we’re doing our best to keep up! One of the obstacles is the divide between experienced ML engineers and engineers looking to make insights about their data. We’re combating this with probabilistic programming, which is the concept of introducing probabilistic primitives as first-class language feature, as well as the inference methods to make them useful. We’ve tied these features together in a developer-friendly way by extending our model of coroutines.In this talk, I’ll highlight some of the ways in which we’ve extended coroutines to enable probabilistic programming, including:The continuation model in HackSource-to-source transformCoroutine suspension in arbitrary expression positionsThe multi-shot modelRepresenting a function as a distributionBasic inference methods over probabilistic coroutines