iiinova — where your digital self keeps evolving

Anton Strezhnev @astrezh.bsky.social · 8h

But yeah - I think the obsession with different named estimators loses the point. What matters is the set of assumptions you’re willing to defend.

020

Anton Strezhnev @astrezh.bsky.social · 8h

Disagree - I think it’s useful to separate what we actually need for causal identification with whatever method we use to estimate and treat those assumptions separately. Researchers may be willing to defend parallel trends, but effect homogeneity rarely is substantively motivated.

100

Reposted by Anton Strezhnev

James @gravitysra1nbow.bsky.social · 13h

We’re already helping to fund the Russian war effort and we’ve basically ended all other aid to Ukraine. Why the fuck should the Ukrainians care about our threats anymore.

321493308

Anton Strezhnev @astrezh.bsky.social · 10h

Less “BLUE” and more “the F-test on the restrictions is usually not significant” All the new methods are also just TWFE with more parameters (and occasionally IPW weights)

261

Anton Strezhnev @astrezh.bsky.social · 10h

If anything - all of the new methods are just versions of TWFE anyway! Just specify the TWFE correctly and any concerns about TWFE disappear :)

230

Anton Strezhnev @astrezh.bsky.social · 10h

The issue here is that it’s not “apples-to-apples” - saying that you can fix static TWFE bias due to time heterogeneity by fitting the dynamic is already acknowledging that we needed a different estimator.

130

Anton Strezhnev @astrezh.bsky.social · 09/10/2026

Oh it’s like custom rFactor2 lol - even they don’t want to touch the EA games

010

Anton Strezhnev @astrezh.bsky.social · 09/10/2026

Wait, they don’t have the ABS/TC assists enabled? Are they running the F1 game or like an Assetto Corsa mod?

110

Anton Strezhnev @astrezh.bsky.social · 07/10/2026

Okay no new Frisch-Waugh-Lovell results, we’re safe still.

1140

Anton Strezhnev @astrezh.bsky.social · 07/10/2026

Yeah, they sold out Madison pretty quickly :/

010

Anton Strezhnev @astrezh.bsky.social · 07/10/2026

At least no mathematician has ever had an extreme reaction to the rapid development of technology.

14783134

Anton Strezhnev @astrezh.bsky.social · 06/10/2026

Again, lot of mask-off between 2016 and 2026.

010

Anton Strezhnev @astrezh.bsky.social · 06/10/2026

Yeah - though less “popularism” and more “deliverism” - I think the response to Trump I took the “left behind” thesis seriously. Less so after Trump II.

110

Anton Strezhnev @astrezh.bsky.social · 06/10/2026

My view is that 2016 was an attempt to rediscover embedded liberalism to stave off the worst and 2024 was the realization that it wouldn’t have mattered anyway.

120

Anton Strezhnev @astrezh.bsky.social · 05/10/2026

That classic plot device, p-hacking

040

Reposted by Anton Strezhnev

rev. howard arson @theophite.bsky.social · 03/10/2026

actually, AI is really good for manipulating any abstraction layer where you know at a conceptual level how it relates to the one you understand, but where the toolkit you have available at your abstraction layer does not have an appropriate off-the-shelf solution.

41749

Anton Strezhnev @astrezh.bsky.social · 03/10/2026

I’ve also started doing more math-anim stuff with graphs and visuals in lecture just to keep things interesting. Normally the added cost of figuring it out in Plotly is too high, but now I can design the “keyframes” and let the LLM figure out transitions.

060

Anton Strezhnev @astrezh.bsky.social · 03/10/2026

They’ve been great for adding custom interactive widgets inside lectures - like this illustration of Bayes Rule with BISG took like 20 or so minutes w/ a fairly barebones prompt - www.antonstrezhnev.com/ps812/slides...

antonstrezhnev.com

BISG: predicting race from surname and county – PoliSci 812

180

Reposted by Anton Strezhnev

brasidas @brasidas.bsky.social · 02/10/2026

Folks, you have to observe B for this to work!

6769

Anton Strezhnev @astrezh.bsky.social · 02/10/2026

Yeah - have seen a lot of fully AI reviews in recent months. I think it’s on the editors to police this! Presumably they’re also reading them?

110

Reposted by Anton Strezhnev

Joe Noonan @joenoonan.se · 30/09/2026

Social science should be more open with the fact that novelty is smaller and smaller and more constructed. In like 1950, you could be like "we don't know how congressional committees work here is my book on congressional committees". Not the case now.

3747

Reposted by Anton Strezhnev

Jeffrey Wooldridge @jmwooldridge.bsky.social · 29/09/2026

In case anyone is interested, here is a new working paper on non-binary treatments with staggered interventions. Essentially, I show under what assumptions it is valid to estimate cohort x time dose-response functions. papers.ssrn.com/sol3/papers....

papers.ssrn.com

A Heterogeneous Slopes Approach to Difference-in-Differences with Non-Binary Treatments

I extend the approach of Wooldridge (2023, 2025) to allow for heterogeneous gains to treatment in the context of staggered interventions for non-binary treatmen

211135

Anton Strezhnev @astrezh.bsky.social · 29/09/2026

Miss the days when seeing an MP5 was a guarantee of at least a B+ level action movie.

041

Anton Strezhnev @astrezh.bsky.social · 26/09/2026

Never been a better time to work in Terminal. Told my Windows-using students that there’s literally no reason not to install Linux instead. They’ve solved gaming!

020

Anton Strezhnev @astrezh.bsky.social · 24/09/2026

But I guess in your case it’s just a classifier and you just multiply the weights by -1?

000

Anton Strezhnev @astrezh.bsky.social · 24/09/2026

As far as how sensitive the internal representations are to directed post-training - defer to someone with more expertise. My stats intuition is that it’s hard to move a max trained on a lot of data with just a few new data points but things may get weird when you can move a lot of parameters.

130

Anton Strezhnev @astrezh.bsky.social · 24/09/2026

I think the common confusion is that the model has no “reward function” at runtime - it’s always weird seeing the internal representations talking about goals/rewards since the “reward” was just part of whatever gradient descent step moved its weights at training - the cake is a lie essentially.

141

Anton Strezhnev @astrezh.bsky.social · 22/09/2026

Hell is actually reinforcement learning.

050

Reposted by Anton Strezhnev

Ned Resnikoff @resnikoff.bsky.social · 19/09/2026

As I keep saying, professional influencers are petty bourgeois entrepreneurs who make their money by harvesting attention, and that position in the capitalist economy tends to lend itself to a certain kind of politics.

6440102

Anton Strezhnev @astrezh.bsky.social · 17/09/2026

“Moderates?!¯\_(ツ)_/¯”

030

Anton Strezhnev @astrezh.bsky.social · 17/09/2026

Longtermism is an actual Temporal Pincer Movement

060

Anton Strezhnev @astrezh.bsky.social · 17/09/2026

“And having sold all of his people to McCarthy, he then made a film called On the Waterfront which was a celebration of the informer…I have to add, that he is a very good director”

120

Anton Strezhnev @astrezh.bsky.social · 17/09/2026

The accounts going off on “stochastic parrots” are, ironically, very predictable - but there’s a new layer of brainworms every week from the boosters.

110

Anton Strezhnev @astrezh.bsky.social · 17/09/2026

A lot of this might still be account centrality (sorry you’re getting on Discover!). My experience has been mostly good discussion w/o the X/Linkedin hypebeast stuff.

100

Anton Strezhnev @astrezh.bsky.social · 16/09/2026

Of all the days to drop a p(doom) reference when teaching Bayes’ rule

140

Anton Strezhnev @astrezh.bsky.social · 14/09/2026

“Nostalgia is a trap that makes us live in a past that never existed” I say as I hit pre-order on World of Warcraft: Forever.

080

Reposted by Anton Strezhnev

Chris Hayes @chrislhayes.bsky.social · 14/09/2026

This is going to make me sound old and maybe a little nostalgic but it’s really striking how in the Trump era, no one even pretends to debate about policy anymore.

1984785530

Reposted by Anton Strezhnev

Boo!-per Lund @cooperlund.online · 12/09/2026

If I may - I think this is a symptom of YouTube creator culture being a cargo cult where people just adopt the form of whatever seems to please the algorithm. It’s why everyone makes That Face.

0613

Anton Strezhnev @astrezh.bsky.social · 12/09/2026

Need to teach Matt about collider bias

1130