I did a talk at Simons Foundation on how Stan was initially developed and how it got to where it is today:
It was the first time I’d presented this and I was having so much fun I ran out of time near the end. If you’d like to go more slowly, here’s
- slides for the talk [.pdf]
Also, here’s a direct link to the the TV clip that was played near the start of the talk.
- Stan and Julia mentioned on Billions [Twitter]
Bob:
Thanks for posting this. It’s soooo satisfying to be part of a collaborative project that gets used by so many people. I’d like to thank the U.S. taxpayers, along with various other funders, for their support for this and related projects.
One of the fundamental reasons I cite for Stan’s success in the talk is the marketing reach of this blog (and secondarily your textbooks). For marketing purposes, I think Twitter would’ve worked just as well. Of course, I also mention that my reasoning is all entirely retrospective, so who knows if I’ve identified the correct reasons—I’m just trying to lay out what we were trying to do and how we went about it and what I think worked and didn’t. For example, I say that getting governance sorted out was huge (and I mention some missteps along the way). It required me letting go of the reins, which reduced my and Daniel Lee’s stress level(s) tremendously, so I may just be citing things that were important to us more than the project as a whole. Anyway, take it all with the grain of salt as the retrospective commentary it is.
P.S. While we’re at it, let’s also thank the Slovenian taxpayers. Erik Štrumbelj was awarded a Slovenian Science Foundation grant early in the project’s life to build GPU capability for Stan. Erik then brought in amazing grad students like and Rok Çesnovar and Tadej Ciglarič, who subsequently made core upgrades to our C++ infrastructure well beyond the GPU layer. Not to mention Novartis, their financial contributions, and letting Sebastian Weber work on the project—he added core multiprocessing and multithreading that we wouldn’t have otherwise.
As a complete non-expert when it comes to the heavy math/CS side of things, I thought the history lesson of what the technical landscape was prior to Stan was very interesting. I was also idly curious what about R made you mention it as a bad role model as a language/community? Is that because it has really limited governance?
First, thanks for posting slides in addition to the video! I wish everyone did that.
For me, the most interesting part was about project governance and engaging with the community. I think the Stan community is amazing.
I find it surprising that you didn’t do code review initially. I now try to do code review in every project with more than one person. Or if I am fixing an issue in an open source library, try to engage the person who submitted the issue in code review. Of course with C++ that’s a bit more tricky.
Very cool. Thanks so much for Stan! I use it every day (either Stan or a high-level interface like brms).
From my perspective, the high level interfaces like brms and great model checking and viz tools like the loo package and bayesplot have been key to promoting Stan to the masses. On the Stan forum, a large volume of questions involve using those high-level interfaces like brms. These seem to be most frequently from non-statisticians who are using it in their field. The visualization tools and model checking tools have really made a pretty cohesive, complete workflow, possible (and relatively user friendly). Like you say in the slides, the community is great.