iiinova — where your digital self keeps evolving

Rex "garbage in" Douglass @rexdouglass.bsky.social · 25/01/2026

Worst Day So Far SITREP - Authoritarian Consolidation Last Update Jan 24, 7:11 PM · LLM Dec 1, 2025–Jan 23, 2026 (54d searched)

172

Rex "garbage in" Douglass @rexdouglass.bsky.social · 24/01/2026

Worst Day So Far SITREP - Authoritarian Consolidation Last Update Jan 23, 11:22 PM LLM Dec 1, 2025–Jan 23, 2026 (54d searched)

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 22/01/2026

Today is the public launch of Worst Day So Far I built it to solve a simple problem that I faced, and I suspect many of you face as well. Please check it out and let me know what you think, and what would make it more useful to you.

140

Rex "garbage in" Douglass @rexdouglass.bsky.social · 22/01/2026

An authoritarians strategy is to spam horrible things to saturate your attention and critical thinking. I built this tool to leverage AI to fight back. Who else needs this tool? @meidastouch.com @atrupar.com @ronfilipkowski.bsky.social @briantylercohen.bsky.social @acyn.bsky.social

010

Rex "garbage in" Douglass @rexdouglass.bsky.social · 22/01/2026

WorstDaySoFar.com Daily summary and tracking of US mass rights violations and democratic collapse Disappearances, warrantless detentions, dragnets, political prosecutions, and 250+ other types of events Fully automated and community funded

161

Rex "garbage in" Douglass @rexdouglass.bsky.social · 02/10/2025

Embedded the first frame with an encoder then arranged in 2d with UMAP

010

Rex "garbage in" Douglass @rexdouglass.bsky.social · 29/09/2025

👋 Hiring teams working on 𝗗𝗼𝗰𝘂𝗺𝗲𝗻𝘁/𝗩𝗶𝗱𝗲𝗼 𝗔𝗜, 𝗜𝗘, 𝗼𝗿 𝗶𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗮𝗹 𝗖𝗩—I’m open to full‑time or freelance. DM or rexdouglass@gmail.com

010

Rex "garbage in" Douglass @rexdouglass.bsky.social · 29/09/2025

𝗪𝗵𝘆 𝗶𝘁 𝗺𝗮𝘁𝘁𝗲𝗿𝘀: surfaces the weird, rare, and important—so labeling targets 𝘀𝗵𝗼𝘄 𝘂𝗽 𝗮𝘂𝘁𝗼𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆 and you get 𝗯𝗮𝗹𝗮𝗻𝗰𝗲𝗱 𝘁𝗿𝗮𝗶𝗻/𝘃𝗮𝗹/𝘁𝗲𝘀𝘁 𝘀𝗽𝗹𝗶𝘁𝘀 without scrubbing terabytes by hand. 𝗦𝘁𝗮𝗰𝗸: fastdup · UMAP · H.264 metadata 𝗨𝘀𝗲 𝗰𝗮𝘀𝗲𝘀: anomaly mining, active learning, dataset curation for industrial vision.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 29/09/2025

𝗪𝗵𝗮𝘁 𝗜 𝗯𝘂𝗶𝗹𝘁 (𝗥𝗜𝗢𝗦 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝘁 𝗠𝗮𝗰𝗵𝗶𝗻𝗲𝘀): 1. 𝗞𝗲𝘆𝗳𝗿𝗮𝗺𝗲𝘀 from 15s clips 2. 𝗦𝘁𝗿𝗮𝘁𝗶𝗳𝗶𝗲𝗱 𝘀𝗮𝗺𝗽𝗹𝗶𝗻𝗴 over H.264 stats (oversample rare quantiles) 3. 𝗦𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀 + 𝗰𝗹𝘂𝘀𝘁𝗲𝗿𝗶𝗻𝗴 (fastdup) 4. 𝗥𝗲𝗽𝗿𝗲𝘀𝗲𝗻𝘁𝗮𝘁𝗶𝘃𝗲 𝗽𝗶𝗰𝗸𝘀 per cluster → high‑signal frames 5. 𝗨𝗠𝗔𝗣 visualizations to audit coverage & outliers

120

Rex "garbage in" Douglass @rexdouglass.bsky.social · 29/09/2025

🟢 𝗢𝗽𝗲𝗻 𝘁𝗼 𝗪𝗼𝗿𝗸 — 𝗔𝗽𝗽𝗹𝗶𝗲𝗱 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 (𝗥𝗲𝗺𝗼𝘁𝗲 / 𝗔𝘂𝘀𝘁𝗶𝗻) → rexdouglass.com 𝗛𝗼𝘄 𝗱𝗼 𝘆𝗼𝘂 𝗳𝗶𝗻𝗱 𝗮 𝟭‑𝗶𝗻‑𝟭,𝟬𝟬𝟬,𝟬𝟬𝟬 𝗲𝘃𝗲𝗻𝘁 𝗶𝗻 𝟭𝟬,𝟬𝟬𝟬+ 𝗵𝗼𝘂𝗿𝘀 𝗼𝗳 𝘃𝗶𝗱𝗲𝗼? You 𝘀𝘁𝗿𝗮𝘁𝗶𝗳𝘆 𝘁𝗵𝗲 𝗵𝗮𝘆𝘀𝘁𝗮𝗰𝗸. #ComputerVision #ActiveLearning #DatasetCuration #MLOps #VideoAnalytics #Manufacturing #AppliedScience

240

Reposted by Rex "garbage in" Douglass

Rex "garbage in" Douglass @rexdouglass.bsky.social · 26/09/2025

🟢 𝗢𝗽𝗲𝗻 𝘁𝗼 𝗪𝗼𝗿𝗸 — 𝗔𝗽𝗽𝗹𝗶𝗲𝗱 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 · Remote / Austin 🎞️ 𝗦𝗹𝗶𝗱𝗲𝘀 tinyurl.com/yxr7k5f8 🚧 𝗣𝗿𝗼𝗷𝗲𝗰𝘁 Completely unsupervised segmentation + labeling of parts, tools, and people in manufacturing. I demo it on a random “How It’s Made” fire-extinguisher video. 🧰 𝗦𝘁𝗮𝗰𝗸 DINO · SAM-2 · Gemini · YOLO

131

Rex "garbage in" Douglass @rexdouglass.bsky.social · 26/09/2025

📜 𝗕𝗮𝗰𝗸𝘀𝘁𝗼𝗿𝘆 This is the POC I built to land an offer at RIOS. I inferred their internal pipeline and automated the whole thing—end-to-end in ~10 days. The work is paused while VC funding sorts itself out, but it was a blast and we were on track to productionize and scale quickly.

020

Rex "garbage in" Douglass @rexdouglass.bsky.social · 26/09/2025

🧠 𝗠𝗲𝘁𝗵𝗼𝗱 LLM drafts a scene narrative LLM proposes label vocabularies Grounding DINO generates bounding boxes LLM filters boxes SAM-2 propagates high-confidence masklets Distilled YOLO runs at the edge (Definitions—scene, frame, object instance, masklet—standardize the units.)

120

Rex "garbage in" Douglass @rexdouglass.bsky.social · 26/09/2025

131

Reposted by Rex "garbage in" Douglass

Rex "garbage in" Douglass @rexdouglass.bsky.social · 24/09/2025

Help! ─ 𝗔𝗽𝗽𝗹𝗶𝗲𝗱 𝗦𝗰𝗶𝗲𝗻𝘁𝗶𝘀𝘁 open to work ─ Remote/Austin → www.rexdouglass.com With new tools, I've been cooking: ▸ Machine-vision pipelines — RIOS Intelligent Machines ▸ Information-extraction pipelines — Microsoft ▸ Interactive SMS surveys — Pantheon Insights

196

Rex "garbage in" Douglass @rexdouglass.bsky.social · 24/09/2025

Rex W. Douglas PhD Applied Scientist (Remote/Austin) Looking for full time and freelance projects. Hoping for somewhere stable. I've never been more productive in my life, but mass layoffs and funding collapses have been endemic.

112

Rex "garbage in" Douglass @rexdouglass.bsky.social · 24/09/2025

196

Rex "garbage in" Douglass @rexdouglass.bsky.social · 24/09/2025

000

Rex "garbage in" Douglass @rexdouglass.bsky.social · 15/09/2025

Some ML Engineer at YouTube must handle just kids home for the holidays unsubscribing their parents from all the right wing channels.

040

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/09/2025

Follow for more pro tips about destroying old hard drives last minute for a move.

020

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/09/2025

90% of machine learning is data cleaning.

192

Rex "garbage in" Douglass @rexdouglass.bsky.social · 06/09/2025

Who turned America's military against Americans? I'm tracking 313+ instances of support/opposition to the use of armed forces in domestic law enforcement in a new dataset: "Public Positions on Militarization of Domestic Law Enforcement in the U.S." docs.google.com/spreadsheets...

docs.google.com

Public Positions on Militarization of Domestic Law Enforcement in the U.S.

050

Rex "garbage in" Douglass @rexdouglass.bsky.social · 28/08/2025

Public Positions on Militarization of Domestic Law Enforcement in the U.S. t.co/P5HmsHF1fA

010

Rex "garbage in" Douglass @rexdouglass.bsky.social · 15/08/2025

No clue, there are lots of other parameters to the bargaining model of war even before you get to the domestic political constraints. I'm just very annoyed that there's taboo about seriously measuring the costs, and finally got to it.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/08/2025

Just the direct material costs for each additional year of war to Ukraine are about $111.7B or just over half their entire GDP burned per year 🚩$50B (25% of prewar GDP) to wartime defense spending over peace time levels 🚩$61.7B (30.8%) in direct destruction of things/sectors of the economy by Russia

140

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/08/2025

This is absolutely nuts: 🚩30% of Ukraine's population is either internally displaced, a refugee outside of the country, or under Russian occupation. 🚩About 14% of the population's homes have been destroyed. Each additional year of war destroys about another 3%.

050

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/08/2025

Each additional year of war in Ukraine implies an additional: 7,953 Civilians Directly Killed 12,822 Civilians Directly Injured

061

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/08/2025

The bargaining model of war only has a few parameters. I'm often annoyed that the cost of fighting receives less serious measurement than military weapons/strategy. Then I remembered I pay $700 a month for LLMs, so here's the first in a series on Ukraine's costs from fighting.

070

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/07/2025

Please reach out if you ever want to talk shop about the intersection of -manufacturing -machine vision -generative AI -and automation/robotics

000

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/05/2025

I still have a number of those projects brewing including executing large scale automated literature review with LLMs. You can see an early alpha of that work here. rexdouglass.com/Natural%20La...

020

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/05/2025

I can't say much more than that so I'll point you at my public facing work prior to going private. I have 27 public facing projects at my website. Among them are large scale natural language processing projects, spanning both LLMs and teams of a dozen people. rexdouglass.com

130

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/05/2025

-What I built worked. We could automate big manual processes, halving the manual effort/time, usually beating the humans. My projects worked because I insisted on modeling the actual data generating process and identifying (1) what LLMs could do and (2) what only experts could.

120

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/05/2025

-Whatever the cutting edge stack in generative AI space was, I was on it. Spent a lot of time developing methods to get things to scale. It's dangerously easy to make a demo and almost impossible to make a production system that works every time. LLM code has a half-life of weeks

120

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/05/2025

-Customers loved me and I loved customers. A social science background made me better positioned than I expected for corporate work. Day 1, I'd interview everyone who had ever touched that data, have a full game tree / code book mapped, know their workflow better than they did.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/05/2025

Still navigating NDA land but things I can say for sure: -I love applied work. I was building models directly for internal customers and watching it succeed or fail in real time in weeks instead of years. This is a vastly different cadence than academia and much more fun.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 14/05/2025

Based in San Diego and ideally looking for either remote or local work (Austin also a possibility with local family). Open to anything data facing, from applied science, to analytics, to ML engineering. Based on some of my repos, folks are trying to even move me to a SWE.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

If that's actually your research question, then god be with you, but more likely what you have done is thrown away all of the interesting variation in your data and you are now modeling effectively noise. What's left over could literally just be the measurement error of the thermometer.

050

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

So instead of asking why San Antonio was hot in 1990, or hot relative to other cities, or relatively hot for San Antonio, I'm now asking "Why was 1990 a more extreme year for San Antonio than 1990 was an extreme year for Anchorage?"

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

What I have mechanically done is shift the dependent variable from that city's temperature to how much more/less that city deviated from its own average relative to the deviation from the average of other cities for that year. This idiosyncratic noise is what you're left with to model.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

Now say my research design is confounded by both things about the city I won't bother to measure (like latitude) and things about years I won't measure. So I go, "aha!", I'll just use two-way fixed effects and include both year and unit dummies.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

What I have mechanically done is shift the dependent variable from that city's temperature, to that city's temperature relative to itself over time. I'm now asking whether that year was particularly cold or warm for that city.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

Say my research design is confounded by city properties I don't understand or can't measure, e.g. I don't know their latitude. So I include (only) a unit fixed effect to stand in for latitude and everything else I won't measure about the city.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

What I have mechanically done is shift the dependent variable from that city's temperature, to that city's temperature relative to other cities. I'm now asking what makes this city warmer or colder than other cities, not what makes a city warm or cold.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

Say my research design is confounded by annual events I don't understand and can't measure well, so I include a time fixed effect to pretend to "control" for those unmeasured confounders.

110

Rex "garbage in" Douglass @rexdouglass.bsky.social · 13/04/2025

Recently had a debate about two-way fixed effects and how it pretends to magically solve causal inference problems by just changing to a different DV than the one we actually care about. Here's that explained visually. Here's raw annual city temperatures:

181

Reposted by Rex "garbage in" Douglass

Posit @posit.co · 03/03/2025

We are delighted to announce Reticulate 1.41! You can now simply declare your dependencies using `py_require()`, and Reticulate will handle the rest. This seamless experience is powered by uv, an extremely fast #Python package manager written in Rust. Learn more: posit.co/blog/reticul... #RStats

012229