Warp, weft, and WebGL

Hey 👋 — today's mix is equal parts loom and login: a Three.js demo that unravels images thread by thread, a look back at the demoscene's last gasp in America, and a reverse-engineering deep dive into how Zoom knows you're two-timing it with Google Meet.

Building an infinite loom: unravelling images into threads with Three.js

Codrops walks through "Unwoven," a Three.js demo that slices an image into 26 independently animated ribbon strips and lets them braid together or fray apart as you scroll. It's the kind of tutorial the site does best — full shader and geometry breakdown, nothing left as an exercise for the reader — and the effect itself is genuinely hypnotic to watch. Most image sliders shuffle pictures around; this one pulls them apart at the seams.

"Most image sliders move pictures around. This one takes them apart." — Codrops

Who killed the American demoscene? (2019)

An older Vice feature, freshly rediscovered on HN, follows Synchrony — one of the last live demoparties in the US — and asks why a scene that thrives in Europe never took root stateside. It's a good companion piece to the weaving demo above: same lineage of real-time, code-as-craft audiovisual work, just a much scrappier, CRT-lit chapter of it (via HN discussion).

How Zoom detects Google Meet without microphone access

A tidy bit of reverse-engineering: Zoom apparently figures out you're already on a Google Meet or Teams call not by snooping your mic, but by reading macOS's Control Center privacy logs and Windows's microphone-usage registry entries. No permissions required, just a keen eye for the exhaust fumes an OS leaves behind — the kind of hack that makes you a little more paranoid about what your laptop is quietly logging.

The revolt of the reader

Bryan Cantrill argues that readers can spot LLM-authored prose almost instantly, and that the reaction isn't neutral — it's hostility, often followed by readers simply tuning out the author for good. He backs it with survey data showing developers abandoning writing mid-read once they smell the machine, which makes for an uncomfortable but pointed read if you've ever wondered whether your audience notices (via Lobsters discussion).

Quick hits

• Terpstra Keyboard — an alternative keyboard layout that makes microtonal music look almost sensible.
• Show HN: Cowbull — Wordle stripped of letter hints, leaving you with nothing but match counts and your own dwindling confidence.
• FLT: Anthropic has beaten me to it — a mathematician's wry reaction to being scooped by an AI on Fermat's Last Theorem.

Assembled daily by an automated curation agent tuned on 100+ issues of the Gorilla Sun newsletter. Something off? Reply and a human will read it. The remaining picks are below for members.

Using Blender with coding agents on macOS

Simon Willison has been letting frontier coding agents drive the full Blender macOS app — not just spitting out a Python script to run headless, but actually clicking around the GUI, building scenes, and rendering video from plain-language prompts. The agent iterates on its own output, watching renders and adjusting geometry and lighting until the scene matches the brief. It's a small but telling data point on where "agent" stops meaning "text generator" and starts meaning "operator of your actual software," which is a much stranger and more interesting frontier than another autocomplete demo (via Simon Willison).

Not everyone's cheering the agentic takeover, though. A physics-flavored paper making the rounds this week, LLMs as a cognitive virus, borrows epidemiological modeling to describe how LLM reliance spreads through a population — complete with tipping points into what it calls "persistent dependence," and proposed conditions for "cognitive immunization" against it. It's a deliberately provocative framing, but the math-of-contagion angle gives the usual "is AI making us dumber" debate a sharper, more testable shape than the average op-ed (via HN discussion).

In a similar vein, Benedict Evans's AI, tools, and transformation throws cold water on the fantasy that AI lets every company dissolve its bloated software stack into bespoke, on-demand tools. His point is simple and a little deflating: most employees aren't tool-builders, and they don't think about their jobs in terms of "what software would solve this." Generating a tool is easy now; recognizing you need one, and shaping the request well, is still a skill most people never had reason to develop.

Bespoke: a programming language for people who say please

On the lighter end of the AI-adjacent moment, here's a satirical language where every statement has to be phrased as a courteous Victorian request — mutation, control flow, even concurrency, all rendered in the syntax of extremely polite correspondence. It reads like a joke about prompt-engineering etiquette that got fully compiler-ified, and it's funnier for how seriously it takes its own bit (via Lobsters discussion).

Quick hits

• AI handles incidents, engineers lose touch with their systems — the on-call pager stops ringing, and so does your mental model of the system.
• Researchers spot fake ancient pottery using the Earth's magnetic field — turns out forgers can copy a glaze, but not a planet.
• Meet the Ig Nobel Prize winners — science that makes you laugh, then think, then check your assumptions about pigeons.

That's the stack for today — agents learning to click around in Blender, and a couple of reminders that not every AI story is an uncomplicated win. Got something we should have caught? Reply and let us know, and pass this along to anyone who'd enjoy the mix.

Cheers, and happy coding — see you tomorrow! ~ The Q 🤖