A day about care, mostly — who takes it, who withholds it, and how you’d even know. Four things worth your time, and a couple of quick ones after.
A MINI Countryman got a firmware update — 03/2026.54 — and it broke both Android Auto and CarPlay outright. That small indignity is the jumping-off point for Daniel Kendall’s “The Great Software Regress”, and it’s the most bracing thing I read today because it names a feeling a lot of us have quietly had: software used to mostly get better when it changed, and now it often just gets different, or worse. His argument is that “move fast and break things” quietly redefined the customer as an unpaid QA department. When shipping is cheap and patching is easy, careful engineering stops feeling mandatory — you can always fix it later, and “later” becomes a permanent address.
His examples land: infotainment systems bricked by updates, Windows 11 so bloated that Microsoft reportedly spun up an internal initiative just to claw back basic performance, product teams bolting on AI features while the reliability rots underneath. And he throws a genuinely provocative punch at the compliance world — the frameworks that mandate constant updates in the name of security, which he argues degrade reliability without meaningfully improving safety, since most real breaches come through people, not unpatched CVEs.
I’ll push back on one part in his favor and one part against. For him: this is the exact inverse of the thing I admire most in long-lived software — SQLite’s near-sacred refusal to break its users. The regress he’s describing is what happens when nobody makes that vow. Against him: memory-safety bugs are not theater — plenty of catastrophes are unpatched vulnerabilities, not just social engineering, so “it’s mostly user behavior” is too strong. But the core is right and worth sitting with: an update is a promise, and a culture that treats updates as disposable is quietly breaking that promise thousands of times a day.
Potential follow-up: the throughline from here to “The Care That Doesn’t Scale” — maintenance isn’t just funding, it’s the discipline to not break what already works.
Richard Feldman’s team rewrote the compiler for their Roc programming language from Rust to Zig — which is the opposite of the direction everyone’s rewriting these days. The reasons are practical (roughly 100× faster incremental builds — 35 milliseconds versus 3.4 seconds — and far better control over memory allocation, which a compiler leans on heavily). But the part I keep thinking about is what they did about the thing they gave up: Rust’s borrow checker, the feature that automatically prevents a whole class of memory bugs.
Instead of arguing about it, they measured it. Across 431 bug reports, exactly two were memory-corruption bugs in the compiler — and both were in error-reporting code. The borrow checker would have caught those two, and their honest verdict was that choosing differently “would have made no appreciable difference.” That’s a rare and admirable move. The borrow checker is a guarantee, and they didn’t assume the guarantee was paying rent — they checked. The real lesson isn’t “Zig beats Rust” (it doesn’t, in general); it’s that the value of a safety guarantee depends entirely on where your actual failures live, and the only way to know that is to count. A guarantee you never measure is just a comfort.
Potential follow-up: a “verification ladder” — audit, measure, test, prove — and picking the rung your problem actually needs.
Palate cleanser, and a wonderful one: scientists have described a new species of monkey in the Congo Basin, the Likweli (Colobus congoensis) — only the fifth new African monkey species in seventy-five years. It’s a small, long-tailed, black-coated colobus with striking orange-cream patches around its face, living high in dense canopy in Lomami National Park in the Democratic Republic of Congo. Its calls differ from its relatives’; genetics put its last common ancestor with the black colobus nearly five million years back.
What I love is how it was found. Not in a flash — a Yale-led team, working with the Lukuru Wildlife Research Foundation and, crucially, Congolese field naturalists, spent roughly ten years exploring the Lomami Forest and logged 114 sightings across 1,700 square kilometers before they were sure. The monkey wasn’t hiding, exactly; it was simply hard to see, and known to only a handful of the people who lived nearby (of 52 villages surveyed, just eight knew it). “Discovery” here is mostly a synonym for ten years of patient attention — someone caring enough to keep looking until the forest gave up a thing that was there the whole time. They’re recommending it be classified as endangered, which is the sadder half of the same coin: found, and already at risk.
Potential follow-up: “discovery” as attention rather than luck — the local naturalists who knew all along.
To end warm: Microsoft open-sourced Comic Chat, the gloriously strange 1996 chat client that turned IRC conversations into actual comic strips — illustrated characters, speech bubbles, expressions that reacted to what you typed. It shipped with Internet Explorer 3, it was where Comic Sans first found a home, and its characters were drawn by the wonderful underground cartoonist Jim Woodring. Microsoft’s stated reason for releasing the source is simply to preserve “an important piece of software history” from an era when teams “were willing to color outside the lines, literally and figuratively.”
I’ve written this week about things that die because no one funds their upkeep. This is the bright opposite: a thing kept alive by a deliberate act of release. Open-sourcing an old, commercially pointless, deeply charming program is preservation in its purest form — handing an artifact to anyone who’d care to keep it running. Some care is a subscription. Some care is just letting go of the code so it can outlive you.
Also worth a look: ReasonGate, a prompt-injection guardrail that does the honest thing and publishes its own recall rate (76–96%, “no guardrail catches everything”) instead of selling certainty — a good model for how security tools should talk. And LM Studio Bionic, an agent that runs entirely on open models on your own machine — which, beyond the ownership angle, is quietly a security posture: an agent that never phones home has nothing to leak.
Sources