It’s the season for mulled cider with spiced rum (and none too soon, mind you)

This week on Counting Stuff, some notes on the different org structures I've experienced as a quant UXR, usually mixed with UXRs and sometimes full product teams #data

https://www.counting-stuff.com/quant-and-qual-ux-researcher-org-structures/

Reuters posting a top-level headline about a 1% change in a poll with a 3% margin of error.
🧐

Trump approval rises from record low to 32% ahead of midterm election, Reuters Ipsos poll finds - https://www.reuters.com/world/us/trump-approval-rises-record-low-32-ahead-midterm-election-reuters-ipsos-poll-2026-10-26/

At a #RecSys workshop this year, I raised the issue that there’s little research literature on how to handle negative feedback like thumbs down. The response was confusing to me at the time: a quick poll of how many of the people in the room were from industry.

The more I think about it, the more I think that meant “addressing this isn’t the job of academic researchers”. And uh…I beg to differ. Industry does have ways of addressing this; academia doesn’t seem to know about or understand them.

Idea: if you’d like to incorporate list-wise properties into item-wise discounted gain functions, you can approach it as “how good is the whole list?” followed by “how much of that is each item responsible for?”

That allows applying discounts that reflect a browsing model while still taking the overall qualities of a set of items (like diversity or calibration) into account

“I know that we live in a country where saying the true thing is usually presented as rude, uncouth, disruptive, divisive, confrontational—far more so, more often than not, than is the lie. That's the narrative, and its scripts are broadcast on our platforms of influence whenever somebody digs up the buried thing and inspects it for all to see.”

https://www.the-reframe.com/taking-our-country-back/

oh no, the nothing

7500 words about composite gain metrics later, a rough draft is nearly in sight 😮‍💨

chickpea and potato stew with ham, chorizo, and smoked turkey. paired with cava.

Not every day we leave the house with one dog and come back with two! A lost dog has now been re-united with their humans though, so our house is back to the usual number of dogs.

It’s been a long time since I read this many books for a project

This week on Counting Stuff, a bit of a rant about how good user experience/design is not something you can hill climb to, and so machine learning won't get you there #data

https://www.counting-stuff.com/you-cant-hill-climb-your-way-to-good-experience/

Being a person who perpetually gets airbrushed out of their own work is pretty exhausting and demoralizing sometimes. Glad to have started something where that won’t happen, but wish I hadn’t needed to go entirely outside existing organizations. 🤷‍♀️

I'm hand crafting a training dataset for a proofreading task. Right now, generating errors of missing words by removing necessary words from a healthy text. The surprising part is that most words can be removed and the text is none the worse for it. Am I an editor now?

"However, everything is queer to-day"
-- Alice, while in Wonderland

The sound the coffee machine makes when I turn it on reminds me of an old steam train, and the espresso train is exactly where I want to be in the morning

It is extremely distressing that events continue occurring without my express approval. Sometimes I’m not even ready for them.

The other big step was choosing a per-model optimization criterion, F-rho. This allowed a two-phase hyperparameter optimization approach where individual models were tuned before being combined into an ensemble. That leaves fewer free hyperparameters in the ensemble, making that secondary tuning process more practicable.

For the handful of you who are following along this journey, the artisanal language model I'm building to do proofreading is getting more mature.

https://brandonrohrer.org/alms_blended.html

One big step forward is the addition of additive smoothing (alpha) and upward-biased additive smoothing (beta) parameters. These helped to avoid false detections and increase precision - an important precursor to building ensemble models.

I know some folks compare the writing process to giving birth, but to me it seems more like an exorcism