there's this ridiculous idea that's conned most of the software industry, that human engineers can and will be replaced by AI. that AI is cheaper than the human labor that it replaces. that driving adoption means tokenmaxxing, that AI vendors will continually reduce costs and improve the models. that the playing field has been leveled and anyone with an AI agent at their disposal now has godlike powers.
these past couple of months, the chickens have started coming home to roost. all the prices have gone up, but the vendors aren't making any money because it's still costing them more compute per token than they're asking for. so they, in turn, are dialing down their models because the dumber they are, the more loops it takes to get stuff done.
very few companies can afford the burn.
my client is one of the many that can't. so, weeks after a town hall in which they explained how finances were tight, followed a short time later by explaining that autonomous agentic workflows are going to be the primary mechanism for software delivery and on-call resolution, followed by a week of costs going up while productivity tanks... i arrived at the office on wednesday morning to discover that everyone had been allocated a $12 per day AI budget.
average costs per day had been in the high double digits to low triple digits the whole week.
there was a massive uproar, and the budget was increased to $30 per day. barely enough to get anything done, and engineering is beginning to understand that this effectively means returning to the good old days of having to understand what you're doing and craft code like peasants. there's only one problem, though: last year the company fired half its staff, and rotated all the rest, so nobody actually has any experience with the software they're officially responsible for. hell, barely anybody even knows which software is under their ownership. at a guess, i'd say that the ratio of software engineers to software repositories is somewhere around 5:100, which is something i think about a lot since learning about conway's law and organizational scaling.
it'd be funny if it wasn't so sad. and the saddest part is that these blunders are hurting real people, with real families to provide for and real mortgages to pay off.
...
i spent a large chunk of my day fixing a deployment pipeline - some random devops guy had made a change a couple of months ago that literally did nothing but break it - while expensively asking the AI tough questions to ensure that it wasn't making shit up. i also had a lot of issues with the company-mandated claude code wreaking havoc with any and every plugin or skill i attempted to work with.
i did take a pleasant break to coach sailor on working with AI, and continuously snipe him for insisting on staying south africa instead of coming home.
...
i came home, took mr smear for a pleasant walk around the park, then returned for dinner, then took another walk with gd and my sister as well, for a tour of our old neighborhood and a good view of a particularly impressive moon.
after getting home and showering, we watched a little but more of redline, then put mr smear to bed and repeated the trick of me taking a puff before passing out and actually sleeping again.
gd says my snore is the most comfortable she's ever heard.
...
i began yesterday doing random shit on the balcony, including trying to help the ex-saffer i met a few weeks ago whose department has just been laid off. i caught up on quite a bit of reading, watching, and had a big breakfast before heading out to the office.
i started my day trying to work with haiku, but anthropic's cheapest, shittiest model is so useless it seemed to be trolling me and i immediately switched back to sonnet. work-wise, i spent most of my day planning and executing a task i had been expecting to get to on monday or tuesday, and thinking very, very hard about my personal AI usage and what a harness should look like. and trying and failing to get anything real done on that last score.
i hit my $30 limit shortly after lunch, but the software to track it hadn't been installed yet, so it seemed fitting that that was the point at which i voluntarily restarted my machine and synced with the rest of the organization :P
...
that was around the time i got a phone call out of the blue from my mentor, who informed me that my client has apparently decided to stop working with outsiders altogether and that the two of us currently embedded with them will be moving on in a month's time. apparently this is a money thing, because their feedback for both of us has been positive, so i guess my town-hall feelings from a few weeks ago weren't wrong...
i asked what i should be focusing on to make sure i get to my next client as quickly and smoothly as possible, and he informed me that between my auto-doc project and consistent publishing of articles he doesn't believe i'll have too much trouble. this is good.
i waited until my counterpart had been informed by his mentor, and the two of us sat on the balcony for half an hour strategizing. he'd been taking the whole thing a bit personally - his new manager has been behaving very strangely since the big reshuffle a couple of weeks ago - and i think it helped him when i put the pieces together and suggested that the manager probably knew this was coming and just isn't very subtle about trying to get ahead of it.
either way, we have a month to focus on up-skilling in AI and we're in a pretty good place to be doing so. i also told him about the up-skilling solution i pitched to the boss, and he's interested in helping me make that happen, so that's good.
...
by the end of the day, i'd installed codex with my personal subscription on my work computer, and delivered a first iteration that looked pretty good, and had an awkward encounter with my client who's pretending everything's cool until they break the news to us on sunday, and had a wild political debate with the AI guys that was actually quite fun.
and stopped myself from "just one last deployment" on a thursday evening :P
mr smear wasn't at home when i arrived, he'd gone off to a new friend's place for a swim. i quickly shared an update with gd (i thought my sister could hear me, so i had to re-tell over dinner), then headed to the hairdresser. crazy singing guy was outside again, harassing the nail salon next door, and i got a pretty decent haircut - i think? - and then came home to shower and do a little more AI work until mr smear got back and we could sit down for onigiri (to celebrate the girls finding a mold) and corn on the cob.
the AI work i did was validating. codex is simply better.
dinner was delicious, then i did a little more fiddling around with AI and unity, and then we all sat down to finish watching redline. redline is a phenomenal movie, with a bizarrely jarring ending (kind of like grease), but really the greatest pleasure in watching it is seeing it as an art delivery mechanism. it's gorgeous, and it's simply mind-blowing seeing how gorgeous it is knowing how it was made.
...
i took another puff last night, but it didn't seem to take. am i already developing a tolerance after two nights? last night was long, and my lower back hurt a bit, and i barely got any sleep between lying restless and uncomfortable, and then i finally had enough just over an hour ago and climbed out of bed and am now sitting on the balcony typing this with half an eye on the world waking up around me.
