Tried tackling two problems at a time -- one agent writes and works while I steer and make decisions about another problem with a different agent. This. Fucking. Sucks. I'm not doing this ever again
I was going to write a blog post about expected loss, expected value of sample information, and expected value of perfect information, but I didn't really have much to say aside from just deriving stuff for my own edification, so here is a markdown file instead github.com/Dpananos/Obs...
github.com
The good news is: STATISTICS is only hard if you want to do it accurately ! 😅😭😂🫠
Chelsea Parlett@chelseaparlett.bsky.social · 2d agoThe good news is: Marketing measurement is only hard if you want to do it accurately ! 😅😭😂🫠
It’s been like 10 years since I was able to whiteboard out stuff with some fellow students. PhD wasn’t the same, I was a man on an island.
Access to my university's library is so clutch. Its like "Try before you buy" -- which is good because holy is this book expensive
Demetri@phdemetri.bsky.social · 6d agoLong weekend plans looking Bayesian to me
On device transcription models are nice (free, on device obv_ but man the cloud providers are just do damn fast. How do they do it?
Canada has this thing called “Rib fest”. Take a guess what this is
A measure theory competition ? I’m sure you’ve got it in Lebesgue
Theory is all fun and games until you have to have someone who knows the theory better than you read your shit
Claude doing god's work. Now if he could just clean my house
More thoughts about the normal prior vs the student's t prior dpananos.github.io/posts/2026-0...
Adaptive Shrinkage and the Student’s t Prior – Demetri Pananos Ph.D
dpananos.github.io
Just shipped Student's t priors for our Bayesian AB testing. The interesting part? No MCMC, no numerical integration. How did he do it!?!
New rule: Statistical analysis without a clearly stated research question gets you a paddlin'
I was planning to wrist a blog post at work but Gelman and crew already wrote it for me in their new book :(
I have a lot to say about Bayesian ab testing all of a sudden.
LinkedIn thinks I want to work for FAANG. In reality I want to do math and walk my dog without worrying about slack.
Doing math with AI does not feel as good as doing math without ai, but more like having a friend (much smarter than you) give you just the hint you need to gain traction. Traction is what matters to me now. Traction pays the bills.
I'm not sure where I got this, probably somewhere on twitter or the R user universe, but I have stopped using "simple" when describing mathematical procedures and instead use "straightforward". I forget the original argument for doing so, but I feel less like a chump
Imagine having a native speaker of a foreign language ask to practice their english with you. Their english is not perfect, but you can impute what they are trying to say and help them improve. Now imagine if they said english words but in a sequence which had no meaning. That is what I'm feeling
Demetri@phdemetri.bsky.social · 2w agoI find it incredibly difficult to comment on ad hoc quantitative approaches. I have nothing to really grasp onto to tell me if things are good or bad.
I find it incredibly difficult to comment on ad hoc quantitative approaches. I have nothing to really grasp onto to tell me if things are good or bad.
So I'm _really_ interested in a specific problem. However, the problem is far beyond my ken: approximation theory, functional analysis, numerical methods which I haven't touched in a decade. Chatgpt is telling me some very convincing stuff, but I just don't want to trust it ...
alcoholic trying to get sober after a relapse tells Google's "Health Coach" about it. it tells them to drink every hour and messages them to make sure they're drinking as soon as they wake up.
I miss meetings of 4-6 engineers figuring out how the fuck we're going to build this crazy new feature. Now we just point claude/codex at the problem. I certainly don't miss the effort, but I do miss the camaraderie.