babysitting

 i'm starting this post at almost half past midnight, and i've just managed to resolve an issue i've been struggling with for hours on end, herding an AI cat to solving a problem that's entirely out of scope but very much interferes with what is in scope.

i've been thinking a lot about context, lately. it's my job, but it's also a new frontier with... ahem... many different ways to skin a cat. i've also been thinking a lot of the AI vampire, the death of the craftsman engineer, and why replacing developers with AI is going horribly wrong. while i've been neck-deep in context engineering rabbit holes and side quests, and pondering how much of this experience might shape my own concept for an IDE.

it took far too long for me to realize that i needed to hack the LLM loop and make it more robust, which i did by instructing it to document the state of the code, then document each attempt to fix it, and append lessons learned to the end of the document to be used for the following iteration.

from there, it took eight iterations, at the end of which i instructed it to add an epilogue to the documentation explaining why it took eight iterations after restarting with this technique. and that epilogue was the first part of the entire process that i was actually able to understand, because it had been looping over and investigating interactions between a bunch of "black boxes" interacting with each other, at one point even going so far as to dive into multi-dependency byte-code.

all of this wrangling, while running a whole bunch of other small code and configuration changes, each of which requiring its own AI babysitting.

i didn't have a headache when i finally got home, but my brain hurt. and i still wasn't done, and i was still bothered by the main obstacle, so i couldn't let it go.

...

ironically, all of the above was done because i didn't have the credentials to run the test scripts i wrote this morning (also using AI, because now there's no reason to write a simple test script when i can just specify sample generators, test scenario generators, test runners, validators and clean-up scripts that all work nicely together).

...

this morning was a big one for gd; she's had a really rough few days (after a rough few months, to say the least), but we got her to the examination having followed all the instructions and procedures, and afterwards i got her follow-up appointment booked (post-biopsy) and got her home safely.

in the evening, i left the apartment almost as soon as i arrived to go looking for a frying pan, because ours is warped and gd's not handling (especially not with everything else going on). that hunt turned out to be a) not fruitful and b) very, very far. so i got back not only with an exhausted mind, but also physically tired. and then i took the recycling out with mr smear, and received an intense (but justified) talking-to about using the neighboring building's recycling bin (because ours has been stolen, twice), which was embarrassing to say the least. and then, just as i got back upstairs, big data sent a photo of some serious harry potter gear for purim so i launched out the door again to walk to the other side of the neighborhood to pick up the bag of goodies.

and then we ate (dinner was delicious), and then i had way more dessert than i should have, and then it was shower and bedtime, and then it was the extra AI babysitting accompanied by loads of youtube videos.

i wonder if i'll sleep tonight?