Darren’s blog was very helpful as I got started learning about bayesian inference, so it’s very neat to see Rainier appear there. (Also, he’s been a wonderful beta tester and now contributor. Thanks!)https://twitter.com/darrenjw/status/1002685644777377799 …
2/3 Stan is obviously extremely useful and successful in its niche, and I think the focus and investment there has been appropriate. The itch I was trying to scratch has to do with lowering the bar for production engineers to work with bayesian models,
-
-
and in that context, *any* non-standard language is huge roadbump. The dependency on the gcc toolchain is also a (not insurmountable, but real) barrier to deployment.
-
I see. So I guess to summarize the Stan perspective (as best I can): functional is not strictly more elegant than imperative, learning a C derivative is the least of a budding statistician's problems, and inference algorithms are probably not meant to be deployed (insights are).
-
My personal views on those issues have definitely shifted over time (not sure where I'm at right now). Deployment is perhaps most arguable - their view stems from a fairly sturdy belief that inference algorithms aren't perfect and diagnostics should be inspected with ~every fit.
-
Again, I think Stan made the right choices for its audience. I agree that it comes down to deployment. My question is, *if* you decided to deploy inference at scale, *then* what would you want? I do think a familiar production environment is part of that.
-
Functional vs imperative isn’t that interesting a question to me here aside from the case where you already have engineers who are used to working in a particular environment and paradigm - at which point being as native to that env as possible is a big win.
-
Like, we could geek out about ways Stan could be a better PL, as a PL, but I think we agree that “being interesting to PL geeks” shouldn’t be one of Stan’s goals. “Being interesting to Scala geeks” *is* one of Rainier’s goals.
-
We have a debate pretty often regarding whether you can entice engineers into learning how to do good statistics by making it closer to something they know, much like the way you're trying to do with Rainier. I think almost everyone disagrees with you and me on this issue :P
-
I have my beliefs here but tend to go with theirs since they've been in that intersection a lot longer than I have. This should be a good experiment!
- 2 more replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.