Tertulia

"Full Self Driving" https://www.youtube.com/shorts/1N6RXQIWaiE

i'm realizing this website's redesign worked because i didn't ask too much of claude code. i have another website i asked it about, one that requires a stronger visual language and... well, it didn't work. that's not to knock the product, of course: last night's exercise caused me to wonder about gemini, copilot, chatgpt, and deepseek. all whiffed way harder than claude, which had good design instincts, but could not actually execute its own design intent. useful tool, wrong task (for now).

holy smokes, it's 02:38 am, gngn 💤

today was an interesting day with claude code. it's gotten a lot better since opus 4.5. i honestly haven't been using it because the regular chatbot experience works well, and i learned on prior versions of claude to more tightly constrain my problem descriptions to get more accurate responses. it seems that may not be necessary anymore.

since february, i have been studying for the jlpt n3 exam. i wanted claude's help to show my progress here.

there's now an about page, with a new feature for tracking my vocabulary progress, with data straight from my anki exports.

the only real trickiness here was preparing my anki deck for being displayed:

1. i had an ancillary deck predating my Ankidrone collection that was throwing off metrics, claude helped me merge it into my existing decks without losing interval data.
2. i was missing words from bunpro, which artificially caused my progress to appear less than it is. claude helped me to merge my bunpro progress into my existing decks, creating new cards for words i'd learned on bunpro but did not exist yet in any deck in my collection.

the real insanity is that rather than writing to the .apkg directly, it analyzed my exported collection offline, then handed me scripts that drive an anki extension's http interface — complete with a dry-run mode so i could analyze its efforts before committing to them.

this was a whole set of, "if i had to do all of this myself, it wouldn't have gotten done." yet it is.

the only other hangups were in ops; claude used a library new to my app's stack — sqlite, for reading anki .apkg files — but immediately identified the issue and patched it. i had to provide some oversight over database migrations.

i think the magic here may be that llms are good at doing stuff they've seen before and nothing here is suuuuper novel. but hey: wow.

follow my jlpt journey at /about!

[Nicole] recognizes the genocide in Gaza, supports an immediate arms embargo, and believes our tax dollars should invest in our communities, not fund war crimes overseas.

the biggest opportunity i see for the government given recent advances in llms is the obvious one: paying off tech debt.

it seems that most federal applications start accumulating tech debt the moment their dependency chain is locked in for the first time. once `pom.xml` specifies a specific version of Spring, you're never going to see that version increment ever again, forever.

i've seen multiple programs now on aging codebases that... well, really just too old to be maintained anymore. applications whose core dependencies are at least a decade from having gone defunct (looking at you, Stripes). i've seen Apache Struts installations as old as my college days. i actually do spend time, not a lot, but more than none, debugging why newer versions of Tomcat won't play nice with these ancient frameworks. that, too, has been made easier by LLMs being quicker to pinpoint breaking changes across a range of version updates that were never incrementally applied way faster than i can.

the problem is that our security controls are updated, our applications break, and because we don't have the time or the budget to fix our applications; or our development teams just can't figure it out; instead we ask for exceptions. and they're typically granted, resulting in an enterprise software architecture that has varying levels of compliance, and exposes a patchwork of exceptions to potentially exploit.

plus it complicates upgrading later: if our custom Stripes has a bunch of custom patches to keep it operational, when it comes time to upgrade later, how do we merge the new library with our own patches? do we still need them? what if we only need some of them because, say, a particular bug was fixed?

thus it seems to me that the first step as an llm-powered federal software engineer isn't to let the llm take over and do everything as it's been on legacy codebases, but to iteratively pay off tech debt and get our apps closer to current.

ok! that was mostly painless. my one-button build/deploy system still works despite not being used since... january. there were a few bugs, but claude code was able to iron those out too. claude code also added light/dark modes (i didn't have those before), and wrote the pagination feature for me. still my own homebrew'd scala 3 / http4s core under the hood. there've been a few milestone releases to http4s v1.0.0, nothing to write home about (or worry about here). haven't hit a 500, or an unusable UI component yet. we good 🕶️

it may be time to actually add pagination to this thing. scrolling goes on forever.

and/or fix the background so that it doesn't scroll off into infinity given an infinite number of posts; simple rotation around the x-axis should suffice. 💭

there are a few bugs, but this platform is really everything i wanted it to be: a quiet little corner of the internet, with a decent rich text editor, where i can respond to the news without a news feed constantly trying to get me upset to keep me on the platform, or other platform features — like facebook's recent foray into bullshit notifications — carefully engineered to separate me from my attention.

screw you, mark.

(i have a colleague building some super saas-y things right now, maybe i should take a cue myself and invest some time into the learning and self-improvement features i said i'd wanted at the outset. i've got a good library of machine learning exercises i've been meaning to do something more with than to simply complete. though i've done that too.)

Dear Stephen Miller,

Cancers starve the host organism by hogging all the resources for themselves.

Like billionaires.

With love,

— Everyone

abusive use of “domestic terrorism” investigative authority is not a new problem, but it is a serious one

www.aclu.org

How NSPM-7 Seeks to Use "Domestic Terrorism" to Target Nonprofits and Activists | ACLU

What President Donald Trump’s latest memorandum targeting civil society does, does not, and cannot do.