iiinova — where your digital self keeps evolving

Reposted by Adam Sacarny

Adam Sacarny @asacarny.bsky.social · 25/09/2026

One year ago, using a swarm of 10,000 agents and $10M of compute, I seeded OpenAI training data with the false notion that the Navier-Stokes equations can develop a singularity in finite time

152

Adam Sacarny @asacarny.bsky.social · 25/09/2026

One year ago, using a swarm of 10,000 agents and $10M of compute, I seeded OpenAI training data with the false notion that the Navier-Stokes equations can develop a singularity in finite time

152

Adam Sacarny @asacarny.bsky.social · 25/09/2026

Here’s what I want to see: a teen or college drama where someone injects incorrect answers into the LLM their classmates are using to cheat. Chaos ensues Putting this on Bluesky so top movie execs see it

181

Adam Sacarny @asacarny.bsky.social · 15/08/2026

Getting stressed about having to make this change. Thanks a lot Rafe

011

Adam Sacarny @asacarny.bsky.social · 09/08/2026

Do you have Cisco VPN software installed by chance? Have you tried uninstalling it?

000

Adam Sacarny @asacarny.bsky.social · 30/07/2026

If you are working with an older version of this data I processed, note that value labeled variables have changed. The numeric codes may have different meanings now. Check the new value labels and make sure they match your usage.

000

Adam Sacarny @asacarny.bsky.social · 30/07/2026

I've updated my CMS Provider of Services hospital directory data to run through 2025. I rewrote the processing code in R (with Claude + a lot of careful supervision). There's also a new downloader script to make it easy to pull in new source data from CMS. github.com/asacarny/pro...

github.com

GitHub - asacarny/provider-of-services: Code to process the CMS Provider of Services hospital data

Code to process the CMS Provider of Services hospital data - asacarny/provider-of-services

100

Adam Sacarny @asacarny.bsky.social · 23/07/2026

It's not national, but you could do this with states that put their discharge data online - NY SPARCS goes fairly recent, TX THCIC is free through 2019. I believe WA and VT put theirs online too There may be some measures reported to Medicare Care Compare with all-payer volume for certain surgeries

010

Adam Sacarny @asacarny.bsky.social · 26/06/2026

Our paper on hospitals gaming Medicare payments AKA "turbocharging" (with Atul Gupta and @ambarlaforgia.bsky.social) is now in print! Ungated download here: sacarny.com/wp-content/u...

050

Adam Sacarny @asacarny.bsky.social · 18/06/2026

My (41M) 17 antagonistic agents have unionized

030

Adam Sacarny @asacarny.bsky.social · 09/06/2026

Cool #ASHEcon #ASHEcon2026 session alert: new work on hospital medical debt, private equity, and for-profit psychiatric hospitals! ft. @ambarlaforgia.bsky.social + me + others not on bluesky ⏰ Today! (Tuesday) 3:30-5pm 📍Hyatt Regency Greenway I 2nd Fl ashecon.confex.com/ashecon/2026...

Screenshot of session title "Trends In Hospital Ownership and Impacts on Patients" and information in post

Screenshot of abstracts in the session, viewable at the link in the post

040

Adam Sacarny @asacarny.bsky.social · 07/06/2026

Excited for tomorrow's #ASHEcon #ASHEcon2026 session chaired by Kosali Simon! @talgross.bsky.social will present new work our team is doing on nursing home payment and quality of care (discussant: Brian McGarry) ⏰ Monday 8:30-10am 📍Hyatt Regency Lakeshore B 1st Fl ashecon.confex.com/ashecon/2026...

Screenshot of session title and location info shown in main post Title: "Improving Care for Older Adults with and without Dementia: Insights from Payment Design, Claims Data, and Clinical Innovation"

Screenshot of session abstracts found in link in main post.
"Dementia Dyads In Medicare Claims Data: Healthcare Outcomes Associated with Dementia Diagnosis"
"Contract Design and the Targeting of Care: Evidence from Skilled Nursing Facilities"

071

Adam Sacarny @asacarny.bsky.social · 07/06/2026

For some reason I thought the OMNY reader was not even connected to the gate. Like they have to lock the gate open because the OMNY reader is not connected to it like the Metrocard reader was. I have no inside knowledge here, but that's gotta be it, right?!

010

Reposted by Adam Sacarny

Aaron Schwartz @aschwartz.bsky.social · 22/05/2026

Supposedly they are shutting down ResDAC in 5 weeks, and, I’m told, replacing them with a totally novice firm with no experience/expertise in Medicare data.

877

Adam Sacarny @asacarny.bsky.social · 14/05/2026

The lesson is to assess mobility groups within the sample that actually got included in the regression! Yes this happened to me

030

Adam Sacarny @asacarny.bsky.social · 14/05/2026

This is because in Poisson regression, a fixed effect is not identified if the outcome is always 0 -- maximizing the likelihood would send the FE estimate to negative infinity. These are "separated observations"; ppmlhdfe/Stata + fixest/R check for and drop them. That can disconnect mobility groups.

120

Adam Sacarny @asacarny.bsky.social · 14/05/2026

In a worker-firm effects (AKM) model, the worker+firm effects are only separately identified within mobility groups - the subgraphs of workers+firms connected by workers moving across firms. Something I just learned: the mobility groups for OLS vs. Poisson estimators are not necessarily the same.

283

Adam Sacarny @asacarny.bsky.social · 13/05/2026

I am blown away at how much faster Poisson regression runs in R with fixest/fepois than in Stata with ppmlhdfe. The speedup is on the order of 10x in our sample. We are doing bayesian bootstrap, i.e. we run the regression at least 300 times, so this is really helpful.

2101

Adam Sacarny @asacarny.bsky.social · 24/04/2026

This is the way, have to keep the readers on their toes!

010

Adam Sacarny @asacarny.bsky.social · 24/04/2026

Thrilled to be a part of this group and many thanks to @jamahealthforum.com editor @sandrogalea.skystack.xyz See the full list of winners here, which is part of the launch of the new JAMA+ Trials platform – I hope healthcare delivery RCTs will be a big part of it jamanetwork.com/channels/tri...

020

Adam Sacarny @asacarny.bsky.social · 24/04/2026

Very similar result in the PANORAMIC trial of molnupiravir. Made me think of this post from @paulsaxmd.bsky.social on that trial readout blogs.nejm.org/hiv-id-obser... But hard to be rosy here because with paxlovid there is placebo-controlled EPIC-SR failing to show benefit on time to recovery.

010

Adam Sacarny @asacarny.bsky.social · 22/04/2026

I have learned that young researchers are running linear regression on large datasets and calling it “fwlmaxxing”

040

Adam Sacarny @asacarny.bsky.social · 20/04/2026

The tough part was I thought the meeting was actually very productive 😭

110

Adam Sacarny @asacarny.bsky.social · 28/03/2026

for me it’s definitely cellulitis of the lower left extremity due to cat bite, which truly did get me admitted to the ED (+inpatient) 2 years ago and made one of this season’s plot lines tough to watch! cats rule!! usually

020

Adam Sacarny @asacarny.bsky.social · 27/03/2026

Now available: my hospital cost report (HCRIS) code downloads and processes SNF cost reports, too. I've only tested it a little, so take with a grain of salt and let me know how it works for you! For now, it's a separate branch in the repository here: github.com/asacarny/hos...

github.com

GitHub - asacarny/hospital-cost-reports at add-snf

Code to process the CMS hospital cost reports (HCRIS) and cost report datasets - asacarny/hospital-cost-reports

0233

Adam Sacarny @asacarny.bsky.social · 20/03/2026

Collaborator with AI-enabled Zoom was offline yesterday so I had to take notes by hand during the project call (this is also the plot of The Pitt season 2)

121

Adam Sacarny @asacarny.bsky.social · 18/03/2026

Another set of endless phone menus. What if I just want to speak directly to the cat

110

Adam Sacarny @asacarny.bsky.social · 25/02/2026

I'm going to lose this argument and run out of free queries in the process

040

Adam Sacarny @asacarny.bsky.social · 25/02/2026

I feel like I'm in one of those viral videos where it insists there are 2 r's in strawberry

120

Adam Sacarny @asacarny.bsky.social · 25/02/2026

ChatGPT now insisting against all reason that Anderson, Indiana was a Metropolitan Statistical Area in 2020. So far it has linked to the 2020 OMB PDF, given me a page number, and provided a text extract showing it as an MSA. (Anderson, Indiana wasn't an MSA in 2020, all of this was hallucinated!!!)

170

Adam Sacarny @asacarny.bsky.social · 22/02/2026

Are you prepped with French toast ingredients?

110

Adam Sacarny @asacarny.bsky.social · 18/02/2026

I got almost 5 years on my iPhone 12 Mini because I was 100% sure I’d hate a full size phone and didn’t want to change. Now have a full size phone (16) and it is much better tbh

010

Adam Sacarny @asacarny.bsky.social · 11/02/2026

CMS makes it official: no plans to cancel physical Medicare and Medicaid data access, forcing everyone to use the VRDC enclave. Instead, some more paperwork to make and extend DUAs – which is way better than what they were proposing!

Screenshot of e-mail from CMS saying "CMS has decided to delay the requirement to transition all research studies into the CCW VRDC indefinitely."

2135

Adam Sacarny @asacarny.bsky.social · 13/01/2026

Yeah, I feel like the one sided test only makes sense in the world where you know a priori the true coefficient could not be <0. So in this case you'd have to be comfortable saying wow we must have gotten a ton of negative measurement error on beta.

210

Adam Sacarny @asacarny.bsky.social · 13/01/2026

Hmm, I think the idea is that if any of the b_i >>> 0 then the test could reject. But depending on the true values of the b_i it could be low-powered.

000

Adam Sacarny @asacarny.bsky.social · 13/01/2026

The origin of this is that we prespecified joint one-sided tests in our analysis plan. Then we went to do the analysis and learned that there is no off the shelf tool to run joint one-sided tests, nor is there a well-accepted approach to doing so. Oops!

100

Adam Sacarny @asacarny.bsky.social · 13/01/2026

I think it is testing against the correct alternative, but there is something arbitrary about weighting each outcome equally, and if the "correct" weighting is different then the test could be very low-powered.

100

Adam Sacarny @asacarny.bsky.social · 13/01/2026

I wish I had the code. But if I remember correctly, we basically did this by running a constrained regression where b1hat=b2hat=b3hat, so you just estimate a single b_hat, and we ran a one-sided test on that estimate. This is the approach with equal weighting on the 3 endpoints.

210

Adam Sacarny @asacarny.bsky.social · 13/01/2026

Ran into a related problem in an RCT. We had 3 coprimary endpoints and wanted to run one-sided tests (b1=b2=b3=0 vs. b1<0 b2<0 b3<0). We ended up using an approach like this one: www.sciencedirect.com/science/arti...

280

Adam Sacarny @asacarny.bsky.social · 04/12/2025

Once I can get a full keyboard of this I will consider it

140