Sign in

Andy Timm

@andytimm.bsky.social
350 followers 495 following 200 posts

Data scientist @ Grow Progress, testing Dem ads. Usually thinking about methods for survey weighting, efficient causal inference, or scalable Bayes. andytimm.github.io

PostsRepliesMedia
Andy Timm @andytimm.bsky.social · 29/09/2026
saw*, bah
010
Andy Timm @andytimm.bsky.social · 29/09/2026
I assume you say, looks like it uses real usage when doing real work, not just lighter stuff with the dot?
211
Andy Timm @andytimm.bsky.social · 22/09/2026
Plz don’t be in the comic people
020
Andy Timm @andytimm.bsky.social · 21/09/2026
So I just did a quick check, but the main gap I notice, looking at the silver bulletin data at least, has Morning Consult and HarrisX meaningfully redder, whereas plenty of political-first firms I know use panels are much bluer. So maybe a slightly better gloss is in fact MR vs. politics firms?
031
Andy Timm @andytimm.bsky.social · 21/09/2026
MSA also looks better (again, still reading, and so it’s not perfect, but less of a shitshow at least)
010
Andy Timm @andytimm.bsky.social · 21/09/2026
typesafe.ai/legal/terms I think same page you had up before? Obviously, IANAL, and I mostly noticed this when my weekend “hey wtf” screenshot didn’t match another of our engineers’ as of this morning in the “License Restrictions” section at least. Still need to read more.
typesafe.ai
Terms of use - TypeSafe AI
TypeSafe AI is an AI lab building machine-native intelligence infrastructure for automation, designed to make decisions within software. Try our first System One Model, Jev, in early access.
200
Andy Timm @andytimm.bsky.social · 21/09/2026
Looks like they unfucked some of this at least now
100
Andy Timm @andytimm.bsky.social · 19/09/2026
Auto-correct absolutely fucking wrecked me like this yesterday
010
Andy Timm @andytimm.bsky.social · 17/09/2026
Believe this is a router/blend thing- x.com/eliebakouch/... Not certain and they’ve edited the screenshotted oage since, but would be pretty weird to have a page documenting it as that for no reason.
x.com
elie (@eliebakouch) on X
update again: it's a router lol https://t.co/5Zb55dJTfJ good catch @JulienBlanchon
210
Andy Timm @andytimm.bsky.social · 17/09/2026
Think it’s a router?
100
Andy Timm @andytimm.bsky.social · 13/09/2026
Happy sampling! Check it out in the release blog post: andytimm.github.io/posts/nutpie...
andytimm.github.io
Bring Your Own Density Kernel with nutpieR – Andy Timm
Faster sampling via per-model specialization
000
Andy Timm @andytimm.bsky.social · 13/09/2026
Great work like Alchemize from @pymc-labs.bsky.social has illustrated that building custom kernels is a nicely shaped coding agent task with a clear validation loop. As a result, supplying your own density kernel is much easier than it sounds these days, as I discuss more in the post below.
130
Andy Timm @andytimm.bsky.social · 13/09/2026
nutpieR now supports bringing your own custom density kernel! By supplying your own log density and full gradient evaluation, you can frequently obtain severalfold speedups for Bayesian models. In a test across 112 models from posteriordb, my median speedup was 4.8x (!).
graph showing speedups from switching from vanilla nutpieR to nutpieR with custom density kernels. Average speed up is ~7.38x, and median is ~4.8x.
130
Andy Timm @andytimm.bsky.social · 10/09/2026
Exciting! This is a bit like if Floating Points dropped a banger Political Analysis paper
000
Andy Timm @andytimm.bsky.social · 03/09/2026
As someone who hit da bricks for greener pastures a while back, good shit: good choice, enjoy the doing something new!
130
Andy Timm @andytimm.bsky.social · 03/09/2026
Now I want to see the bio for the other survey modes
010
Andy Timm @andytimm.bsky.social · 29/08/2026
Would love to hear more bilingual folks on this, it’s fascinating.
040
Andy Timm @andytimm.bsky.social · 29/08/2026
My experience with French is that it’s not an identical experience but it rhymes? Like the cadence/discursive ticks aren’t the exact same, but it recognizably feels like claudish. I’d be curious to see if that matches a more native speaker’s experience- my French is pretty functional but not fluent
020
Andy Timm @andytimm.bsky.social · 29/08/2026
I’ve been meaning to read this for a bit, good to see this come up
020
Andy Timm @andytimm.bsky.social · 07/08/2026
Did a miss a high profile example of this going on? Curious if folks tweeting about this is a subtweet of some recent paper lol
110
Andy Timm @andytimm.bsky.social · 01/08/2026
Guy who hates AIPAC because they assume it’s Analyst Institute
020
Andy Timm @andytimm.bsky.social · 17/07/2026
This is how I learned this was out today, ty ty!
020
Andy Timm @andytimm.bsky.social · 17/07/2026
I genuinely don’t think I’ve gotten a different opinion than this one in asking about the solo work
020
Andy Timm @andytimm.bsky.social · 06/07/2026
Glad Mexico lost and their fans will be out of the World Cup :)
061
Andy Timm @andytimm.bsky.social · 29/06/2026
Can highly recommend this pairing, I’m very much on my bullshit this week
000
Andy Timm @andytimm.bsky.social · 14/06/2026
The extender strikes again. NBA calling in their least honest officials He clearly isn’t actually allowed to be given a flagrant or he’d have been suspended more than once at this point
050
Andy Timm @andytimm.bsky.social · 04/06/2026
Did Games Workshop let you know early the Emperor was returning? Lovely!
010
Andy Timm @andytimm.bsky.social · 11/05/2026
(If you’re at AAPOR this week and want to chat, hi!)
020
Andy Timm @andytimm.bsky.social · 11/05/2026
This implies the existence of the LPM Group, yes? If so, I’m sure they’re sworn enemies, competing for the same work.
110
Andy Timm @andytimm.bsky.social · 01/05/2026
Isn’t that just PNAS? (Sorry had to do it)
020
Andy Timm @andytimm.bsky.social · 30/04/2026
The bigger problem of course, is that the justices are mostly partisan hacks in robes. But at the same time, a better society might have slightly higher standards for basic causal reasoning (or willingness to listen to social scientists/statisticians)!
000
Andy Timm @andytimm.bsky.social · 30/04/2026
Thinking of this basic point from Ansolabehere and Fraga today for no particular reason.
Abstract from Ansolabehere and Fraga articulating why you can’t “control for party” coherently like the Supreme Court judges think
100
Andy Timm @andytimm.bsky.social · 01/04/2026
IIRC, test-retest reliability isn’t great, and the concepts only indirectly map onto psychometrically valid constructs. There’s also some weirdness in emphasizing dichotomizing e.g. E vs. I when the underlying construct has a lot of middle ground most people occupy. (Not my area of expertise, tbc)
120
Andy Timm @andytimm.bsky.social · 01/04/2026
010
Andy Timm @andytimm.bsky.social · 30/03/2026
Millennial ass TF2 meme about using more bayes
051
Andy Timm @andytimm.bsky.social · 24/03/2026
It’s a bit of both from poking around- They have some misleading language about certain questions not leveraging AI responses, but there are a handful which do have synthetic responses. So they’re both filling out quotas + making the weighting work + actually generating some select responses.
120
Andy Timm @andytimm.bsky.social · 24/03/2026
Realizing I'm likely wrong above, apologies. (There's a lot that's odd here!) TL;DR is some questions do have synthetic responses, so I assume this is to provide more "data" on those and flesh out crosstabs.
000
Andy Timm @andytimm.bsky.social · 24/03/2026
I was poking a bit more, I actually think it's to make a handful of the questions/"crosstabs" possible? Sorry for the bad explanation, a bit spoiled for choice on explanations of why this is weird.
120
Andy Timm @andytimm.bsky.social · 24/03/2026
hrm, realizing I'm wrong about the weights stabilization thing: I think it's to make the crosstabs ("crosstabs"?) viable? Ick. I had a Claude rake this and you can get it to converge to their target pretty ok. Deff is huge, but like not shokcing.
010
Andy Timm @andytimm.bsky.social · 24/03/2026
I'm sure I could find more hilariously weird stuff in here, but instead I'm gonna go touch grass. If someone tries to look at how fucked the raw data from pollfish is, or learns other fun facts about their synthetic respondents, I'd be curious to hear about it lol
200
Andy Timm @andytimm.bsky.social · 24/03/2026
I'll note there exist kinda interesting estimators that attempt to combine human + synthetic sample in vaguely plausible ways (e.g. PPI/rePPI), or ways to attempt to make more realistic synthetic respondents (e.g. subPOP); this is ~none of that. This just seems like janky-ass weights stabilization?
statmodeling.stat.columbia.edu
Better and worse ways to mix human and LLM responses in behavioral research (but you still have to figure what you’re measuring) | Statistical Modeling, Causal Inference, and Social Science
110
Andy Timm @andytimm.bsky.social · 24/03/2026
They posted the raw data for this, fascinating: Notes so far: 1. It looks like these responses exist to make the weights converge, given the synth respondents have NA for non-weighting Qs? 2. The MoE excludes the synthetic respondents. I guess that's better than treating them as 1:1 for humans?
docs.google.com
The Public Sentiment Institute - National Poll (March 20th) - March 21 Release
141
Andy Timm @andytimm.bsky.social · 21/03/2026
Thanks to the lovely r-universe project, I’ve been able to provide pre-compiled bindings for most platforms, simplifying installation. None of this would be possible without @aseyboldt.bsky.social / @pymc-labs.bsky.social ‘s great nutpie, thanks for the great sampler: pymc-devs.github.io/nutpie/
pymc-devs.github.io
Nutpie
020
Andy Timm @andytimm.bsky.social · 21/03/2026
Nutpie is ~2x faster on average than the base Stan sampler on tasks in posteriorDB, though I’ve found that’s more like 5x for my more heavily used, more complex models. Bob Carpenter has a great introductory blog post/paper with some PyMC folks introducing the sampler and explaining the speedup.
statmodeling.stat.columbia.edu
Nutpie: state-of-the-art mass matrix adaptation for HMC | Statistical Modeling, Causal Inference, and Social Science
110
Andy Timm @andytimm.bsky.social · 21/03/2026
Bayesian friends- if you’re curious to try out the blazingly fast nutpie sampler in R, I just put together a pretty lightweight package that’ll compile your existing Stan models!
github.com
GitHub - andytimm/nutpieR
Contribute to andytimm/nutpieR development by creating an account on GitHub.
172
Andy Timm @andytimm.bsky.social · 18/03/2026
r-universe is changing whether I intend to open source something with difficult dependencies. Getting a package with rust dependencies working on CRAN is not my idea of a good time, or a good use my energy :)
100
Andy Timm @andytimm.bsky.social · 18/03/2026
I have a joke about bayes factors but nobody really thinks I should tell it
020
Andy Timm @andytimm.bsky.social · 17/03/2026
The repo name oh my lord lmao
021
Andy Timm @andytimm.bsky.social · 16/03/2026
Oh, other Jaynes. Briefly had a moment of “E.T. Jaynes could totally have a wild ass opinion on this, like his views on measure theoretic probability”. I do want to read Julian though, he’s been been on my fin reading list for ages (regardless of how probable I find his views).
120
Andy Timm @andytimm.bsky.social · 16/03/2026
A big part of the way I think about survey weighting is through connections to causal inference. One thing that emerges from that perspective: regularized weights are pretty, pretty good! I really like @shiraamitchell.bsky.social ‘s discussion of this connection here, esp the Johansson reference.
statmodeling.stat.columbia.edu
Survey Statistics: sampling-weighted loss | Statistical Modeling, Causal Inference, and Social Science
040