Hey, Bob. There are lots and lots of things I want to do with Stan beyond the list above! But I understand we can’t do them all at once.
Regarding the items above:
- Yes, you’re right that there are many orders of magnitude to go before Stan is fast enough that I would be satisfied. Right now we have users who just don’t even try to fit models in Stan because they would take forever (weeks, maybe?). So they’re fitting various approximations and simpler models. Let’s say a model of interest would take a week and a half to fit. One order of magnitude takes that down to a day, which is do-able if (a) you only want to fit the model once or twice, not in some sort of big loop, and (b) you know ahead of time what model you want to fit, so no exploration needed. Another order of magnitude takes it to 2.4 hours, which is better but still makes it difficult to do exploration. Another order of magnitude takes it to 2 minutes, which is just about fast enough that the user can act in a nearly interactive way by playing with models and data.
It’ll take awhile to get 3 more orders of magnitude, but in the meantime I’ll take what I can get, which is why all avenues are important, including brute force, approximate models, and approximate algorithms.
-
The reason I’d like to fit the Monster model (from my 1996 paper with Bois and Jiang) in Stan is that I used it as a teaching example, and it’s worked well in that capacity for many different audiences. So it would be wonderful if, after hearing the Monster lecture, students could go home and fit the model themselves in Stan, play around with the model and data, simulate fake data, etc.
-
The birthday model is another popular example, and it’s on the cover of BDA. Again, it will be great if the model can be fit directly in Stan, for a few reasons: First, the model has problems and I’d like to improve it. Second, the model has other applications than birthdays. Third, I know that we’re working on improved computation for Gaussian process models, and this is a great demonstration example.
-
For the Stan cloud thing, I’m very happy with what Rstudio implemented: https://statmodeling.stat.columbia.edu/2018/10/12/stan-on-the-web-for-free-thanks-to-rstudio/
If paying for cycles is an issue, I’d be happy with a demonstration version. This is not a high priority for me but I wouldn’t be surprised if at some point it will happen, given how much software is runnable on the web right now.