Beboppin’…

5 DOWN THE GARDEN PATH In §4 above, we discussed the ways in which different types of biases can be encoded in the corpora used to train large LMs. In §6 below we explore some of the risks and harms that can follow from deploying technology that has learned those biases. In the present section, however, we focus on a different kind of risk: that of misdirected research effort, specifically around the application of LMs to tasks intended to test for natural language understanding (NLU). As the very large Transformer LMs posted striking gains in the state of the art on various benchmarks intended to model meaning-sensitive tasks, and as initiatives like [142] made the models broadly accessible to researchers seeking to apply them, large quantities of research effort turned towards measuring how well BERT and its kin do on both existing and new benchmarks.1° This allocation of research effort brings with it an opportunity cost, on the one hand in terms of time not spent applying meaning capturing approaches to meaning sensitive tasks, and on the other hand in terms of time not spent exploring more effective ways of building technology with datasets of a size that can be carefully curated and available for a broader set of languages [65, 91].

This is my favorite passage from the original paper, because the claim was quite stark. It’s not just “they’re parrots,” which one could retrocon by saying “… and um, even parrots are good at lots of stuff” :)

The claim was quite bluntly that this whole approach would be wasted effort

Doll is normal.

Screenshot from The Good Place. Jason: “Any time I had a problem and I threw a Molotov cocktail… Boom, right away I had a different problem.”

I think Jason from The Good Place would endorse this strategy.

It’s all relative! Compared to US politics (and the daily stresses of parenting, relationships, health, etc.), rolling the dice on worldwide economic upheaval seems pretty chill.

I knew about the 1969 moon landing. And knew that news of the My Lai massacre came out in '69. But I had those events in separate boxes. Occurs to me now that it must have been disorienting to feel proud of (one aspect of) your country, while also deeply shamed by its crimes.

astronaut: *loading a pistol and getting back on the rocket-ship* bluesky’s haunted.

Thanks to the BlueSky “For You” feed, anything I click like on is recommended to @marylowd.bsky.social and vice versa.

It really confuses our tradition of showing each other memes when we’ve all seen the same posts already!

I seem to spend a few days fighting with pharmacies and/or doctors every month.

Most recently: Insurance system requires that the quantity indicate the number of *inhalers*. Pharmacy system requires that the quantity indicate the number of *doses*.

That one took a bunch of phone calls to resolve.

AI has limits, even if many AI people can't see them On Ben Recht's fantastic new book

In this incredibly detailed and gracious review of “The Irrational Decision,” @himself.bsky.social situates the book in the web of intellectual history and the war between the AI enthusiasts and skeptics.

You can also find the whole album here, including lyrics and links to other streaming platforms (Spotify, Apple Music, more soon):

marylowd.com/music/doo-wo...

Mary E. Lowd - Doo-Wop Warlock (Official Lyric Videos) - YouTube Lyric videos for "Doo-Wop Warlock" by Mary E. Lowd1. Back to Town2. Draw the Summoning Circle3. My Little Imp4. I Didn't Choose Your Name5. Big Sister from t...

My favorite class in World of Warcraft is demonologist warlock. I love being surrounded by all my demons.

This album summons that feeling — one demon at a time — while blending it with bright 60s harmonies and a touch of 80s synth.

youtube.com/playlist?lis...

I never quite knew how one was supposed to experience the legendary San Francisco, even in the 90s before the .com boom. I guess if you’re lucky enough to find expensive parking, then you can eat a bowl of clam chowder in a sourdough bread bowl. But apparently the city makes sense to other people!

I grew up in Cupertino and never understood San Francisco. Sure, you go to the Exploratorium, and … then what? Drive around the hilly streets pretending you’re in the opening credits of Full House?

Screenshot of interactive Claude session. Claude says: "Working on it..." and nothing else.

Current mood.

And AltaVista Babel Fish has been there even longer!

The problem with human software developers is that there’s no way to guarantee they’ll write secure code. You just have to *ask them nicely* and hope they do it

Purple dog captioned "dogs if they were purple"

Here is my Claude Skill for bringing some evidence guardrails to doing a health question/evidence review. It has many opinionated design decisions, and I have written a pretty lengthy README to share the thinking behind it: github.com/DrCatHicks/i...

GitHub - andrehuang/research-companion: Strategic research thinking agents for Claude Code — idea evaluation, project triage, and structured brainstorming. Helps you decide which papers to write, not ... Strategic research thinking agents for Claude Code — idea evaluation, project triage, and structured brainstorming. Helps you decide which papers to write, not just how to write them. - andrehuang/...

My PhD student found these Claude skills to support development of research ideas. I haven't tried them yet myself, but they look very useful and very cool!

github.com/andrehuang/r...

Hey, let's not buddy shame! All buddies are beautiful. ❤️🤖❤️

Mary E. Lowd - My Battleaxe (Lyrics) YouTube video by Deep Sky Anchor Press

If you’re trying to go against all laws of aesthetics and combine Diablo with The Beach Boys, you’ve got to have a love song about your battle axe.

“Loved my battle axe from the very first swing—
The upgrade of a lifetime and the darkness felt it ring.”

youtu.be/wDMVVzneUTE?...

Mary E. Lowd - What Class are You Gonna Train (Lyrics) YouTube video by Deep Sky Anchor Press

Okay! Let’s do this.

If you’re gonna make an album that sounds like how it felt to play Diablo back in the ‘90s — but in bright, shiny, cheerful doo-wop form! — then you’ve got to start with the musical question:

What class are you gonna train???

youtu.be/agcbeuY1XAY?...

I have a really sharp undergrad working on a related problem! If an attacker can poison your context somehow, there's a lot of damage they can do, and preliminary evidence with real programmers (well, students) suggests that (unsurprisingly) the attacks go unnoticed.

AI comes along and makes certain kinds of knowledge work feel more like facilitation, collaboration, curation, communication, etc. Skills that have historically been coded feminine and therefore undervalued.

Humanoid Robots Hit a Turning Point as Their Brains Catch Up The architect of the DARPA Robotics Challenge explains how their brains have caught up

Gill Pratt, Chief Scientist at Toyota, ran the DARPA Robot Challenge in 2015. First time walking humanoids were on the world stage. Earlier he ran the "Leg Lab" within the then MIT AI Lab. Read where he thinks we are now in our robot future (spoiler: its nuanced). spectrum.ieee.org/humanoid-rob...

Because the capybara is the GOAT.

www.poetryfoundation.org/poetrymagazi...

Mary E. Lowd - The Night Janitor and Alien Oceans (Lyrics) YouTube video by Deep Sky Anchor Press

“The Night Janitor and Alien Oceans” was really hard to convert to a song. Nothing was working, until I added the refrain:

“Twinkle, twinkle sensor lights
What lifeforms are here tonight?”

Then it all pulled together.

youtu.be/wx-KRLoilp8?...

Rabbit with many wings and eyes and a halo

BEHOLD! The biblically accurate RABBIT!

Claude, grant me the serenity
to accept the things you cannot change;
the prompt to change the things you can;
and wisdom to know when you're totally gaslighting me.