Michael Ekstrand

This. And even “fast-moving” databases like DuckDB are careful not to break code, especially standard SQL that's worked for 40 years. (This is one of the reasons I like Duck so much — I'm confident my stuff will still work, just faster, as they innovate.) https://fagnerbrack.com/learn-sql-once-use-it-for-30-years-9aceb0bdee03

RE: @firefoxwebdevs@mastodon.social

there are no docs for this in mdn or caniuse yet, so not sure on the status of actually being able to use this, but it looks interesting.

is CSS going to match on the actual tag name, or on the tag adjusted by heading offset?

tfw the theme you use for your lab site finally hits 1.0! and it has a modular theme architecture instead of a big repo you keep merging with!

and one of the release highlights is moving from Bootstrap to Tailwind 😔.

Always seems to blow a few students’ minds when we start talking about psychology of preference and the relationship of preference to observable #recsys data, and we get to some questions:

• is the experience of utility in the moment the same as memory of that utility?
• if they are different, is our goal to recommend a movie or book the user will enjoy in the moment, or one they will remember having enjoyed?
• how might that difference affect system design or evaluation?

Things like this (popping GitHub through a malicious VSCode extension) make me seriously consider going back to Emacs. https://www.bleepingcomputer.com/news/security/github-confirms-breach-of-3-800-repos-via-malicious-vscode-extension/

new things are coming! #lenskit #recsys

New at my blog: keeping track of submissions and work outputs, to make it a lot easier to prepare annual / promotion review materials, compute your personal acceptance rate for your favorite conference, and more. https://md.ekstrandom.net/blog/2026/05/logbooks

Going to #RecSys2026? I'll be giving a tutorial, “Continuous Automated Reproducibility: Practical Tools and Workflows for Repeatable and Reproducible Research”. Don't miss it! More details to come :)

TIL that TOML does not allow you to write `null` and this is A Problem for one of my use cases and .

the github includes in-depth discussion of why “just leave the value unset” is not enough, but nope, no `null`.

A discovery / fresh articulation I had in class today: one of the things I love about #recsys is the direct connection between math and human experience. Here's a statistical property, and here's why naïvely using it would turn your recommender into a door-to-door missionary for the Good News of Bananas.

“In contributor poker, you bet on the contributor, not on the contents of their first PR.” — on PRs, community and AI-generated contributions in Zig. https://kristoff.it/blog/contributor-poker-and-ai/

New one today: someone found the folder where I publish pre-release CI builds of LensKit, and reported it to the Drexel bug bounty for IP leak / dependency confusion / reverse engineering risk.

This is an interesting read about how database designs fall over in the face of AI agents: https://arpitbhayani.me/blogs/defensive-databases/

sometimes the low-tech startup is what the moment requires. https://wheelfront.com/this-alberta-startup-sells-no-tech-tractors-for-half-price/

uggh ubuntu why? 26.04 is finally bringing TPM-backed disk encryption, but it requires some weird snap-based kernel thing. instead of, you know, just a normal systemd initrd like debian, fedora, arch, and everyone else use. https://documentation.ubuntu.com/release-notes/26.04/changes-since-previous-interim/#resolute-tpm-fde-limitations

no *you're* writing a type unification / variable reification function for Python.

Pleased to announce my group's new state of the questionable art hyperparameter tuning strategy, HatTune.

We pull a hyperparameter out of a hat. It is a good hyperparameter, you will like it. We try a couple others from the hat, and they are worse. Do not question the hat.

Well this is interesting. Looks a lot like the kind of things 'jj' has been working to bring to version control. https://github.github.com/gh-stack/

Any friends have experience with Microsoft Loops?

It’s been suggested to me as a possible research lab wiki / knowledge base solution (within Drexel’s existing software subscriptions).

current python frustration: the maze of twisty weight storage formats, not all alike. Spent some time researching how to save parameters for LensKit models & pipelines with something other than pickles, and came away somewhat sad about the current state of play.

I think I can make a pretty workable solution on HDF5, though.

https://github.com/lenskit/lkpy/issues/690#issuecomment-4194748854

RE: @baldur@toot.cafe

Yeah. It's… bad, and also software quality has been a long-standing problem with ebs and flows.

Some of the BS in Claude Code reminded me of when I was doing Groovy metaprogramming, and finding the language had 2–3 different subsystems implementing similar capabilities, completely differently. Except worse.

Very thoughtful essay on #AI, science, research software, the hard work of learning, and what the actual meaningful outputs of academic research work are (spoiler: not the jupyter notebooks of data analyses). https://ergosphere.blog/posts/the-machines-are-fine/

Co-sign this — I find it incredibly rude to just paste ChatGPT at people. If I wanted to know what ChatGPT “thought” I would ask it myself. https://stopsloppypasta.ai/

RE: @jezdez@publicidentity.net

Thank you @jezdez@publicidentity.net and the Jazzband team for keeping a lot of things useful for its run!

The Jazzband project I use (and recommend!) the most is probably docopt-ng — it's my favorite CLI parser for in-project scripts. Separately fom Jazzband, the whole Docopt vision seems to be losing steam too, which is understandable (types are good) but also a little sad.

Very interesting summary of different approaches to improve #python / #pydata performance: https://cemrehancavdar.com/2026/03/10/optimization-ladder/

RE: @hynek@mastodon.social

A tool that has become load-bearing in my setup has an incredibly high feature velocity, and… that’s a little concerning when I take time to ponder it, tbh. They’ve also shipped regressions that broke my setup. But it’s also incredibly useful.

relatively unimportant thing that is annoying me today: my (annoyingly expensive) wifi mesh does not seem to support VLAN tagging of guest WiFi connections.

things that drive me up a wall with AI getting shoved into everything:

right-clicking a word flagged for misspelling in the Outlook editor (and possibly other places, idk) no longer presents an option to add it to my dictionary.

the AI is also not smart enough to know this “misspelled” word is the name of one of the e-mail recipients.

RE: @mdekstrand@hci.social

oh look more github outages (seem to be resolved now, but Actions tipped over this morning).

now that python has real multithreading, time to yeet that custom multiprocessing abstraction layer. https://github.com/lenskit/lkpy/pull/1030

Ahh, #recsys… “you bought headphones, would you like more headphones?”

Do you appreciate news? Do you want to help #recsys research? POPROX News is open for subscriptions!

The Platform for OPen Recommendation and Online eXpermentation is an #NSFfunded project to help scientists and journalists explore better ways to select and present news stories. Our MVP is a daily personalized newsletter — if you'd like to receive that, and periodic surveys, sign up at https://user.poprox.ai/enroll?source=ff&subsource=memd

After signup, you'll receive a confirmation email — check spam if needed.

Kicking off #recsys2024 with a keynote from @Riedl@sigmoid.social on human-centered explainable AI.

If a blind student wants to do a CS undergrad (in US) specifically to work in assistive technology, what schools would be particularly good for that? (Feel free to boost for reach.)

@tedted@hachyderm.io Jonathan has addressed the cost structure here — publication costs aren’t tiny. https://medium.com/sigchi/what-benefits-do-traditional-publishers-provide-83b7bd3fcbc0

While I have frustrations with ACM, I don’t think giving the org ‘s finances a massive haircut will address them or improve the state of computing research and education.

We’re still seeking applicants for a research project specialist / software engineer to help build a news #recsys platform to support user-facing research on news recommendation, personalization, etc. Happy to answer questions! Details and application info here: https://www.linkedin.com/jobs/view/3670339186

Please boost for reach!