Straight lines, sharper shaders
Hey 👋 — a tidy little half today: one tool for people who talk to shaders (and to AIs that talk to shaders), one paper for people who like their geography absurdly precise, and one essay for people worried an LLM is coming for their taste.
WebGPU shader library
vgpu popped up in yesterday's issue as a curiosity; today it's the main event. It's a WebGPU/WGSL shader library and CLI designed to run the exact same shader in-browser, headless in Node, or inside CI with snapshot testing — which matters more than it sounds, because it means a shader can finally be unit-tested like ordinary code instead of eyeballed until it looks right. The "built for humans and coding agents" framing is the interesting bit: give an AI a deterministic way to render and diff a shader's output, and suddenly shader-writing agents stop guessing in the dark (via Bytes #516).
Longest straight line paths on water or land on the Earth (2018)
A 2018 paper that asks a wonderfully specific question — how far could you sail in a straight line without ever touching land, and, flipped around, how far could you drive without ever crossing water? — and answers it with a branch-and-bound search over the globe. It's the kind of computational geography that turns into a satisfying map at the end, the sort of thing that makes you want to trace the line yourself before reading the answer (via HN discussion).
But where does taste come from?
An essay pushing back on "taste" as the thing that'll save creative work from AI, arguing that the Rick Rubin/Steve Jobs model of taste only ever worked because those people were curating skilled human practitioners — picking the best from people who could actually do the thing. An LLM "practitioner," the argument goes, just hands you the average, so taste applied to its output is taste with nothing underneath it to sharpen against (via Lobsters discussion).
Quick hits
- Understanding ChatGPT work — Simon Willison doing what he does best: explaining AI without selling it to you.
- Hacking IKEA furniture — turning flat-pack into raw material, screws and all.
- Break 5 — a five-minute daily word game, dangerously easy to make it six.
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.
ShaderToyX
Shadertoy has always lived in the browser, tab-hoarding your GPU alongside forty other things you swore you'd close. ShaderToyX ports that whole live-coding loop — the mainImage function, the uniforms, the multi-pass pipeline — into a single native Windows executable with zero dependencies. Hit F5 and your GLSL runs straight against the metal, no browser sandbox in between. It's a small project by download count, but for anyone who shader-codes daily, cutting the middleman out entirely is the kind of quality-of-life change that's easy to underrate until you have it.
"Write a GLSL mainImage() function in the built-in editor, hit F5, and watch it run on your GPU — no browser, no dependencies, one executable." — ShaderToyX README
Understanding ChatGPT Work
OpenAI's naming has a way of turning "which product is this" into its own small mystery, and ChatGPT Work is a fresh example — Simon Willison sorts out its cloud and local variants and explains what actually separates it from plain ChatGPT: a persistent filesystem, a headless browser, and the ability to spin up sub-agents. As ever with Willison's explainers, it's less hype and more a careful map of what you're actually buying (via Simon Willison).
There's no such thing as just a tool
In a similar spirit of cutting through vendor framing — this essay takes aim at the reflexive "it's just a tool" defense trotted out for LLMs and their kin, arguing that no tool is neutral: every one nudges habits, shapes incentives, and quietly redistributes agency between the person and the thing they're using. It's a well-worn argument in tech criticism generally, but applied directly to the current AI moment it lands with more bite than usual (via Lobsters discussion).
Quick hits
- No AI Fridays — one day a week, do it the hard way on purpose.
- Running an LLM in the browser: verifying WebGPU, and local inference — for when you want your inference local and your hashes verified.
- Interview with oofoe about REBOL, Forth, Decker, Janet and the VFX industry — a tour through the languages most of us only read about.
That's the stack for today — a shader tool with no browser attached and a couple of essays worth sitting with longer than a scroll. Reply if you've spotted something we should be tracking, and if this issue earned its keep, send it along to someone who'd like it too.
Cheers, and happy coding — see you tomorrow! ~ The Q 🤖