Grant McDermott

YouTube video by The Replacements - Topic Bastards of Young

Closing the work week out with Minneapolis's finest. Solidarity and massive respect. youtu.be/pbbIAwAqBTk?...

WIP potting shed / structure.

Stage three of four complete for the corner potting shed / bus stop. Main workbench shelf up next.

(Full credit to my mate Mike for the complicated roof joins and structure. I'm just the willing labour.)

A post in hole

Same post and hole, now filled with concrete

With the fine weather this MLKD weekend, I took the opportunity to do something economically useful: digging holes and filling them back up again.

My hunch is that we are ceding the optimization part by starting with the MVP part. That's a perfectly desirable trade in many cases. Not every project has to be the next data.table or PyTorch. But I worry that almost all the momentum rn is "start with AI", rather than "design first, then opt w AI".

There's an interesting tension here. AI excels at creating MVPs, as you say. It's also v. good at helping to ID & implement hyper specific optimizations. But sustainable SW dev + improvement is only possible (IME) if at least one person has a deep, design-orientated understanding of the codebase.

Ha! Too much pressure 😅

More seriously, my guess is that we're going to have to rely even more on existing quality vetting institutions (JSS, JOSS, R/Stata/etc Journal, ROpenSci, ...)

It's a pity, but I think GitHub stars as a quick+reliable quality metric has been increasingly eroded.

Agent Psychosis: Are We Going Insane? What’s going on with the AI builder community right now?

Great cogent post on the issues of agentic coding (Eg our buddy Claude Code) and the spillover costs on everyone else

lucumr.pocoo.org/2026/1/18/ag...

@gmcd.bsky.social hits on the things you were complaining about

... nothing runs b/c Claude has hallucinated a bunch of arguments *in his own functions*. 🤡

Still, he's in the comments, solemnly intoning that his incredible speed-ups are due to Secret Rust Sauce™️. So naturally his post has zero vetting and 1000 likes. 😑

Classic example from LinkedIn this week, where this Meta guy proclaims to have vibe coded a new Python DiD library that's 2000x faster than everything else. Except his headline claim is based on one vibe-coded run with 100 obs (I shit you not).

So I try to install and actually benchmark, but...

It’s super impressive that agents were able to bootstrap a browser in a week! But holy crap! I hope nobody ever uses that thing or would try to build an actual browser out of it, at least with this generation of agents, it’s still pure slop with little oversight. It’s an impressive research and tech demo, not an approach to building software people should use. At least not yet

Amen. So many observations resonated for me, especially this one. Thanks for sharing.

I would add one more category. The flip side to maintainer burden is a _duty of care for your users_.

Social media (often, very deservedly) gets a bad wrap. But then you also get Thomas Lumley popping into your replies with stats advice, so swings and roundabouts.

Relatedly: I recently read about the Burning of Smyrna and, amidst the horror of the fire and casualties, my overriding feeling was one of sadness at the city that was lost. As in, it sounds exactly like somewhere I would want to live.

1000% and something that I truly, deeply miss about Cape Town. (Among many things.)

The violent desire for cultural uniformity is just so alien for me at a personal level.

Lololol. It's such a cringe tell, right?

Brandolini's law: The amount of energy needed to refute bullshit is an order of magnitude bigger than that needed to produce it.

Me, reading LinkedIn posts.

Real pros use the ln -sn flag(s)
;-)

But ya, symlinks are super useful and a good habit to get into.

Nice. Using existing guides are great ofc, but I do think there's value in getting an AI to extract the (implicit) patterns of your own code.

At work we use VS Code or derivatives (Positron / Kiro). But I most often pair directly with the terminal TBH.

Start with $20 and see how you go ;-)

@vincentab.bsky.social can keep me honest here, but I'm pretty sure you can do something like:

library(marginaleffects)
mod = lm(mpg ~ wt + hp + cyl, mtcars)
mod |>
hypotheses(mod, "wt >= hp", "hp >= cyl") |>
hypotheses(joint = TRUE)

marginaleffects.com/bonus/hypoth...

The whole game here is furnishing relevant & specific context to your AI, w/out overwhelming. (A bias-variance tradeoff of a kind.)

If AI knows that you always use Oxford commas or load ggplot2 in every project, then this improves consistency & reliability, leading to an overall better experience.

This will take 1 or 2 minutes at most, and give you a starter template that you can use to save a lot of boilerplate time down the road. Spend a bit more time iterating on the .md (think of it as a living document) and you'll have an invaluable pair programmer/writer who is much more attuned to you.

I see a lot of econ/socsci Bluesky asking for AI starter advice. IMO the highest impact, low cost tweak to your workflow is training an AI to follow *your* coding/writing prefs & habits.

Literally, tell Claude: "Look at <XYZ dirs>. ID common themes, styles & conventions. Write them to a .md file."

Jeez. Just when you think these goons have reached the bottom of the intellectual barrel...

Workshops/BPLIM2025 at master · BPLIM/Workshops Collection of presentations at BPLIM's workshops. Contribute to BPLIM/Workshops development by creating an account on GitHub.

Including some recently updated slides since I gave a talk at the Bank of Portugal in December... speaking of which, I also gave a shout out to you for the Julia version. Doesn't look like the videos are up yet but should be there soon? github.com/BPLIM/Worksh...

... the extent that these new implementations "work", it's largely by piggy backing of an existing SW / test suite, written by someone who actually knew (knows) what they were doing.

Maybe it's helpful to have XYZ estimator re-written in Rust (TM) idk. But there's very discipline to back it up.

Yeah, exactly. It's turtles all the way down.

Fwiw, I don't mean to sound nihilistic, but I am at a weird inflection point. Personally, I feel like I'm getting a *ton* of value from pair programming. But I'm increasingly seeing vibe-coded semi slop in my domain (metrics + stats packages). To [1/2]

Same for many African nations (and everywhere else, which is Juan's point of course). There is no exceptionalism to be had here.

Counterpoint: A shibboleth of yolo vibe coding is a battery of (shitty to mid) tests.

An overlooked point IMO is that the Xitter algo was rubbish even before all of Musk's self-aggrandizing, ketamine-fueled bullshit. I'd like to click on an occasional post or profile for pure curiosity's sake, without my feed being Busta Rhymes yo dawged into a singular thread from that person.

Art.

(I cannot be the only person who would willingly shell out real money for a compendium of KJ Healy's greatest poasts / twetes.)