I'd love to hear some things that are unique to this PPL compared to the somewhat crowded space of Pyro/Stan/Edward/PyMC3/Turing.jl/Tensorflow probability.
It seems the innovation here is that the dependency structure of the model is explicit which allows the algorithms to do interesting things like block updates, beneficial for correlated variables. They also have Newtonian Monte Carlo, an exotic, new technique.<p>Oddly though, Stan is still faster in their own benchmarks, despite their statements. In other words, while Bean Machine has a lower wall time, Stan produces more effective samples per second than Bean Machine (and this, not wall time alone, is the important metric).
Can you say more about what’s interesting here versus JAGS for example? JAGS is fully declarative, mature and useful. It has benefits (like being able to work out how best to sample automatically) but it also has huge amounts of expressiveness drawbacks . Stan inherits the BUGS/JAGS syntax but extends it and adopts an imperative language.