Yeah. Thats a good question. I can’t speak for the Turing devs, but I suspect setting default samplers may not have a large benefit and may have the drawback of making the sampler more opaque. NUTS is reasonable continuous variables, but discrete parameters can be challenging and more problem-specific. In my experience, Metropolis-Hastings has poor scalability and can fail in many cases. I suspect it is a default that is overridden in many cases. However, most samplers have reasonable default settings when possible. For example, NUTS uses the same defaults in Stan. Sometimes you can change the target acceptance rate to improve performance. Even with NUTS, the most challenging part can be finding the right transformations to enable efficient sampling. It would be nice if Turing selected an optimal transformation, but I suspect that would be difficult to do 
2 Likes