Gen.Jl

Gen.jl's defining choice is architectural: instead of coupling an inference engine to language compiler internals, it exposes a programmable API — an open-ended set of inference and learning operations including automatic differentiation, but going well past AD into the operations model-based inference actually needs. That is why you can write a novel sampler in ordinary Julia and run it without forking the compiler. Two capabilities stand out. First, custom reversible jump and involutive MCMC, which let you do efficient inference on generative models with stochastic structure and dynamic computation graphs — variable-dimension problems where fixed-dimension samplers struggle. Second, hybrid inference: neural network inference is fast but drifts on out-of-distribution data and needs expensive retraining, while model-based inference costs more compute but needs no retraining and can be more accurate. Gen lets you combine them, taking the strengths of both. The performance story is equally pragmatic: start in the easy modeling language for generative models, inference models, variational families, and proposals, then migrate only the hot paths to specialized modeling languages that generate faster code, and hand-code whatever else demands it. The honest caveats: it is a Julia-only ecosystem today (the team states it is working on ports to other languages), and the performance escape hatch means your fastest code may live outside your main model file. It is a research instrument maintained by a named core team — Marco Cusumano-Towner, Alex Lew, Tan Zhi-Xuan, George Matheos, McCoy Becker, Feras Saad, with Vikash Mansinghka leading the MIT Probabilistic Computing Project — plus open-source contributors. If you use it, cite the PLDI '19 paper. Where it fits: graduate teaching, methodology papers, uncertainty quantification where the sampler is the point. Where it doesn't: beginners wanting plug-and-play inference, Python-first teams, or anyone who wants a hosted service to manage for them.