Yeah, I've been a big fan of probabilistic programming for a while. The real problem is that getting Monte Carlo methods to converge and produce a large sample

Yeah, I've been a big fan of probabilistic programming for a while. The real problem is that getting Monte Carlo methods to converge and produce a large sample from the posterior takes orders of magnitude more time than running an optimizer to descend a gradient. Hey, you can even make it a probabilistic gradient: variational inference! But then you still have a hard time with discrete, nondifferentiable structure.