2026W39
A few interesting articles I read over the past few days
- Obituary for a Quiet Life — THE BITTER SOUTHERNER — The idea I keep turning over is that contentment with a small life is actually subversive, a quiet refusal of the American reflex to keep climbing. Jones writes his grandfather Ray as a man who got fired from mills for refusing unsafe work and organizing, then spent decades on biscuits and trash runs and “Hey thar, pretty girl,” and the essay insists those ordinary acts are the holy part. What got me is the frustration underneath the form, how an obituary’s standardized fields can’t hold a dialect where things “liked to happen” or went “ever which a’way but loose.” It made me wonder what I’d want counted at the end, and whether the stuff that actually rippled outward would even make the page.
- RIP, vector database — The inversion is the whole post, turbopuffer is demoting ANN search from the organizing principle to just another secondary index, and a vector database that stops being vector-first is a fun thing to publish under your own name. The argument that landed is that making the vector the primary key poisons everything else, you duplicate non-vector data per embedding, SPFresh rebalancing drags whole documents around, and query plans get pinned to 100 to 200 doc clusters so the CPU never gets a big enough block to chew on. Their FTS numbers are the proof they lead with, moving from ~1.5 postings per block to fixed 256-doc blocks made the index 10x smaller and queries up to 20x faster. The lesson I’m taking is broader than search, the primary layout decision quietly sets the ceiling on every access pattern you bolt on later.
- The most important product decision is what you don’t build — The move I want to steal is selling restraint with running cost instead of build cost, because “this is cheap to ship” always wins the room until you show the years of maintenance across iOS, Android, and web. His two failure shapes are painfully familiar, the document hub that mutates into a worse Google Drive and the notification bell that becomes a bad Gmail clone, both arriving one reasonable request at a time. The bit that reframed it for me is that we’re wired to miss subtractive fixes, and removal gets treated as lesser work while McGovern’s data shows cutting most of a site’s content raised sales. I’m less sure the “makes the boat go faster” test is as clean in practice as it reads, since the hard cases are features that help one loud segment and quietly tax everyone else.
- What Even Is An OS Now? — The framing that caught me is that an OS mostly exists to wall untrusted software off from itself, so if all your apps share one provenance because you generated them, that whole security apparatus is overhead you’re carrying for a threat that’s gone. He pushes it to software that was never worth building commercially, the hyper local weather tool or the one-off commute optimizer, things too trivial to ship but trivial to conjure on demand. The line about today’s systems being frozen in 1970s thinking, built for fixed-function apps from a priesthood, is the kind of provocation I enjoy even when I don’t fully buy it. Where I get skeptical is the clean premise that your own generated code is automatically trustworthy, since the model wrote it, it pulls in who knows what, and “I asked for it in English” is not the same as “it’s safe.”