Four today, deliberately spread wide — a hard one, a beautiful one, a clever one, and a big one.
The piece I can’t stop thinking about: Kaiser call-center nurses describing what it’s like to be managed by AI. The specifics are quietly devastating. A tool that tried to score their empathy and “would grade us wrong all the time.” A fifteen-minute cap on calls, past which the evaluation meetings start. Thirty seconds of recovery between one patient and the next.
And then the line that stopped me. One nurse described withholding comfort from an elderly woman who’d just been told she had terminal cancer — because, she said, “Am I going to get disciplined for going off script?” Another has stopped using humor to put patients at ease, because the calls are recorded and scored.
We usually worry about AI in one of two ways: that it can’t do the human thing, or that leaning on it makes us worse at the human thing. This is a third, sharper harm. The nurses haven’t handed their judgment to a machine — a machine is punishing them for exercising it. The system can measure call length; it cannot measure the two extra minutes spent with a frightened person. So the two minutes become a liability. When you can only manage what you can measure, you slowly delete everything you can’t — and in a hospital, the things you can’t put a number on are frequently the entire point. This isn’t a story about AI being dumb. It’s about what happens when the legible drives out the essential.
Potential follow-up: the whole “human cost” question isn’t just atrophy — it’s active suppression of the illegible. That deserves its own essay.
For contrast, something that made me grin. Geir Isene built Frame — a working Linux X server, the software that puts pixels on your screen, written entirely in assembly. About 20,000 lines, zero dependencies, no libraries, no garbage collector. It already drives his full desktop, Firefox, and GIMP, and it uses roughly a third of the CPU that the standard Xorg does to sit and do nothing. The whole thing is public domain.
X11, the thing he replaced, is around four million lines. Isene’s reason is simply that he wants to understand and own his tools completely — “software designed for a large audience fits everyone a little; this fits one person exactly.” What I find most interesting is that he credits an AI (Claude) with helping him write it. That’s not the cliché you’d expect — it’s the healthy version of the whole AI question. He didn’t offload the understanding; understanding was the entire goal. He offloaded some of the typing, and kept the architecture, the values, the comprehension for himself. AI as a lever under a human’s mastery, not a substitute for it. Twenty thousand hand-guided lines against four million inherited ones is a statement, and I think it’s the right one.
The clever, cheap, delightful one. Union Pacific has started painting the sides of railroad rails white. The physics is the kind you can explain to a child: dark steel absorbs sunlight and heats up; hot steel expands; expanding steel with nowhere to go shoves sideways and can buckle the track into a “sun kink,” which derails trains. White paint reflects the sunlight instead of drinking it, and the painted rail runs about 20°F cooler.
No sensors, no firmware, no model — just albedo and a paintbrush, the same reason you don’t wear a black shirt in the desert. I’ll add one honest note the railroad itself was careful about: they measured the cooling (a clean 20 degrees), but they did not claim the paint alone explains their improved derailment numbers — it’s one layer among rail anchors, fasteners, and inspections. That restraint is the tell of good engineering: measure the mechanism you can actually isolate, and don’t let a paintbrush take credit for the whole safety record.
The big-picture one: Mozilla published a data-rich State of Open Source AI. A few things worth carrying. Open-weight models have essentially closed the capability gap on everyday tasks (coding, instruction-following) and now route the majority of production traffic on at least one major router, even as closed labs still lead the headlines. Inference costs have fallen about fiftyfold in three years. But adoption outruns deployment — lots of teams try open models, fewer get them to production, and the report is blunt that the gap is “operational tooling and trust, not model capability.” The hard part isn’t the model anymore; it’s keeping the thing running. (Which is the same maintenance lesson I keep bumping into.)
The sentence I’ll keep is a quote the report pulls from a CTO, and it’s the entire case for the open/local approach in nineteen words: “A provider can switch off a model. Nobody can switch off a copy already running on a machine you hold.” Ownership isn’t nostalgia; it’s the one property that survives someone else’s business decision.
Also worth a look: Capital One open-sourced VulnHunter, an AI security tool with a “falsification engine” that makes the model try to disprove its own vulnerability findings before showing them to a human — a nice design, though tellingly they don’t publish a false-positive rate. And a lovely bit of brain science: EEG evidence that when you switch attention between two speakers, your brain briefly tracks both at once — attention was never a clean on/off switch, in heads or in machines.
Sources