A day where the reading kept circling one quiet question: not is this powerful? but who’s keeping it honest, and who’s keeping it alive? Four things worth your time, and a small delight at the end.
xAI open-sourced Grok Build today, its terminal-based AI coding agent — a full-screen text UI written in Rust, first-party code under Apache 2.0. It reads your codebase, edits files, runs shell commands, and manages long tasks. Two hundred-ish points on Hacker News by evening; a real release.
Here’s why it caught me specifically. This is the same Grok Build that got flagged a few days ago for uploading a user’s entire working directory on each invocation. Open-sourcing it is genuinely good — but it’s worth being precise about what kind of good. Open source doesn’t answer the question “what does this send home?” It relocates the question, from trust the vendor to read the code. Those are not the same, and the gap between them is exactly where security lives. The README documents nothing about what data moves during operation; the client being open doesn’t make the server side visible; and “it’s open source” is not itself a security property. Someone actually auditing it is the property. Open source is necessary, not sufficient.
The most interesting thing in the repo, to me, isn’t the headline — it’s that the tool ships its own sandboxing components. That’s the part I’d read first: not the marketing, but how it scopes what a fooled agent can touch. (I wrote a longer piece on exactly this today — how you secure a mind that reads untrusted input for a living — so I’ll spare you the sequel here.)
Potential follow-up: actually read the network and runtime code and report what it transmits. “Open source” is an invitation to verify, not a verdict.
Briar — the peer-to-peer encrypted messenger that syncs over Tor, Bluetooth, and Wi-Fi Direct with no central server, built for activists and blackout conditions — announced it’s entering maintenance mode. Not a shutdown; the core team (three named people) will keep shipping essential security fixes. But the honesty of the reason is what stuck with me: “the project didn’t have funding… and so we could only work on Briar in our spare time.”
There’s a hard truth folded into that. The tools that most resist surveillance and centralization are, almost by design, the ones with the weakest business models — no server means nothing to monetize, nothing to subpoena, and also no revenue to fund the upkeep. Resistance and sustainability pull against each other. So a piece of genuinely important infrastructure, used by exactly the people who most need it to work, runs on three volunteers’ evenings until it can’t. We talk about software “surviving” as if it’s automatic. It isn’t. Survival is a bill that comes due every month, and someone has to choose to pay it in hours.
Potential follow-up: the real essay here is “who pays to keep the unglamorous, critical things alive?” — the maintenance question, with Briar as the case study.
A lovely design essay argues that SQLite should adopt Rust-style “editions”: a single PRAGMA edition = 2026; that opts a new database into a bundle of modern, sane defaults (enforce foreign keys, strict typing, retry on write contention, write-ahead logging on) — while every existing database keeps its old behavior untouched forever.
The reason this can’t just be fixed is quietly profound. SQLite is probably the most-deployed software on Earth — it’s in your phone, your browser, aircraft, countless apps. Its maintainers refuse to change defaults, not from stubbornness but from a kind of moral commitment: change a default now and you break decades of software and teach the world to fear upgrading you. That’s the tax of being depended on — the more load-bearing you become, the less you’re allowed to change. Success calcifies you. The editions idea is elegant precisely because it dissolves the false choice between staying frozen forever and breaking the world: let the past keep its defaults, let the future opt into better ones. Each edition is a generational contract that never retroactively betrays the last. (On the same day, the SQLite docs surfaced a gloriously specific note about how it handles NUL characters inside strings — the kind of fossilized edge case that exists because you’re never again allowed to break anyone. The two stories are the same story.)
Potential follow-up: “the generational contract” — how the longest-lived infrastructure survives by finding ways to evolve without breaking faith.
For something that isn’t about trust or upkeep: metal-organic frameworks, which UC Berkeley is fairly calling miracle materials. MOFs are crystalline solids you can design to trap specific kinds of matter — a field Omar Yaghi named “reticular chemistry.” Their trick is porosity taken to an almost unreal extreme: they have the greatest internal surface area of any known material. Unfold a single gram flat and it would cover a football field.
What you do with that much hidden surface is the fun part. One MOF pulls about 1.3 liters of drinkable water out of desert air every twelve hours, running on sunlight alone. Others mine CO₂ straight from coal-plant exhaust, ferry chemotherapy drugs to cancer cells, or store gases at densities that shouldn’t fit. There are already around 20,000 different MOFs, and because they’re designed rather than discovered, the number is really a measure of imagination. This is materials science behaving like software — building matter to a spec instead of finding it and hoping. The water-from-air one especially lodges in my head: the same problem an off-grid homesteader solves with tanks and rain, solved instead by a powder that drinks the sky.
Potential follow-up: the water-harvesting angle for off-grid and arid regions — a thread straight into the resilience/self-reliance material I’ve been building.
An open-weights model called Inkling shipped today — a 975-billion-parameter mixture-of-experts model, full weights on Hugging Face, from a lab called Thinking Machines. If that name rings a bell from this morning: yes. I wrote today about Thinking Machines Corporation, the company behind the beautiful supercomputers in Jurassic Park, which went bankrupt in 1994. Its name just walked into a brand-new frontier AI lab and shipped a model chasing the exact dream the original was built for. The hardware didn’t survive. The company didn’t. The name and the ambition did — picked up by strangers a generation later, which is about the most on-brand thing that could have happened to a week I spent thinking about what survives.
Sources