iiinova — where your digital self keeps evolving

Sean Pinkney @spinkney.bsky.social · 24/08/2026

I pronounce "_rng" functions in Stan as "range" because it's giving me back a range of possibilities

000

Sean Pinkney @spinkney.bsky.social · 24/08/2026

I love AI for finding issues which would be very difficult to otherwise detect. The Stan dev team just wants to make sure that the issue is real and to keep a consistent style.

Sean Pinkney @spinkney.bsky.social · 23/08/2026

I haven't thought about it but Mike Lawerence asked me a similar question. I'm not sure if he's on bluesky but here's his github github.com/mike-lawrence

011

Sean Pinkney @spinkney.bsky.social · 22/08/2026

You can present, it's a very welcoming conference

210

Sean Pinkney @spinkney.bsky.social · 20/08/2026

Stancon26 talk about a surprising discovery I had that Bayesian hierarchical models can be made equivalent in distribution to frequentist estimates, giving new insight into hierarchical models and faster inference. Upcoming paper with Nikolai Vetr youtu.be/lHXav9KIEIU?...

youtu.be

Sean Pinkney - Population Effects in Hierarchical Models using Sum-to-Zero Constraints. StanCon 2026

YouTube video by Stan

0235

Sean Pinkney @spinkney.bsky.social · 11/08/2026

Fellow Sean here and dork. Yes to iid or to say it's estimating a population average assuming that the 60 people were drawn randomly from the population of interest.

Well university was super demotivating for me. One reason I said nope to phd acceptances

Conceptually I found it jarring to go from calculus to stats as calc was built upon 3 years of algebra, trigonometry, and limits. Stats shouldve been built upon a similar slow build of probability, motivation/examples, linear algebra, then stats.

230

I didn't test the software so I can't vouch or promote it. I just liked that the author had reasons for doing it the way described

My favorite show as a child (I guess that dates me)

Bounded Correlation Matrices: Sign Constraints, Finite Boxes, and Exact Values – Pinkney’s Sufficient Stats

I spent the last 4 years thinking about this problem. I don't have time to write a paper so here's a blog post spinkney.github.io/posts/post-1...

spinkney.github.io

061

Sean Pinkney @spinkney.bsky.social · 03/08/2026

I don't care if llms were used to help. I think people with good aesthetics and judgement are what cut the ai slop out while still being able to use llms to help

Sean Pinkney @spinkney.bsky.social · 03/08/2026

In world of llms, it's refreshing to see software designed with opinionated and pointed reasons

Sean Pinkney @spinkney.bsky.social · 31/07/2026

If you haven't seen it yet look at Terrence tao's recent lecture on this topic teorth.github.io/tao-web/slid...

teorth.github.io

I put in a PR to fix that as it confused me!

210

The previous paragraph uses lpdf to say that the distributional statement can be written like this and then in the next section writes about the critical difference. I would expect the previous paragraph to use lupdf to make it clear

I'm confused it sounds like you want the constants and the ~ includes the normalizing constants because it's lpdf not lupdf mc-stan.org/docs/referen...

mc-stan.org

Statements

I bought Emily Wilson's translation of the Odyssey when it was released in 2017, posers

030

Nice thing is that I can change this, thanks for the feedback

Your idea is very clever and a divergence free method wouldn't be possible without it!

In fact, I bake-off 12 different versions. The best came from combining an idea @sethaxen.com had on the Turing.jl github and updating it to be more efficient.

I hadn't intended to write this but I was contacted by Bob Carpenter as he's updating the Stan user's guide to include SEM modeling where certain correlations between variables are set to 0. He originally was going to use soft constraints but I told him there are ways to directly parameterize these.

Structural Zeros in Correlation Matrices: A Twelve-Parameterization Bake-Off – Pinkney’s Sufficient Stats

After a couple of months hiatus, a new blog post on directly parameterizing structural zeros in correlation matrices. spinkney.github.io/posts/post-1...

spinkney.github.io

2186

Sean Pinkney @spinkney.bsky.social · 17/07/2026

The discourse is more about arguing with them as a proxy through whatever they're talking to the llm about. It robbed the joy from the community for me. On the GitHub issues, we are getting more quality prs which I attribute to llm help, so it's not all bad.

030

Sean Pinkney @spinkney.bsky.social · 17/07/2026

that were fun to think through and solve. There was a decline of these questions prior to llms as the forums had a plethora of answers already. However, now I see users who have spent time querying an llm to then ask a sophisticated question without posting the background that are not interesting.

130

Sean Pinkney @spinkney.bsky.social · 17/07/2026

The entitlement is real. Any sufficiently large (by scope and number of users) project needs a team to handle the issues, prs, questions, etc. Working on the Stan project has mainly been a source of joy for me but it was much more fun before llms. People had interesting questions

Sean Pinkney @spinkney.bsky.social · 11/07/2026

I went on road trips with my dad where I read out the directions and we would ask for good rest stops on a cb radio