--- title: JPEG XL Doubts, Half-Past Fetch, and Tiny GPU Lexers date: 2026-09-15T18:00:00.000+02:00 week_number: 38 notes: "**TL;DR:** This week questions JPEG XL's fit for the Web, looks past `await fetch()` into the still-streaming response, and finds a tiny model highlighting code on the GPU. Plus new HTML experiments, local accessibility checks, Interop 2027 proposals, native secret storage, and local-LLM break-even maths." links: - title: Submit Your Proposals for Interop 2027 url: https://web.dev/blog/interop-2027-proposals comment: Interop 2027 proposals are open through September 23. Strong candidates need a stable specification, no outstanding browser-vendor objections, and solid tests; previous projects have typically already shipped in at least one browser. tag: news - title: The Case Against JPEG XL url: https://giannirosato.com/blog/post/case-against-jxl/ comment: Firefox now plans to enable JPEG XL by default in version 158, but this sharp counterargument says the Web mostly needs lossy efficiency and fast, predictable decoding. Its benchmarks favor AVIF, while a 1.9 KB pathological file keeps jxl-rs busy for 17 seconds. tag: article - title: Half Past Fetch url: https://blog.gaborkoos.com/posts/2026-09-08-Half-Past-Fetch/ comment: "`await fetch()` resolves when the final response headers arrive, not when the body finishes. The article follows that gap through unread connections, stream-teed clones, post-resolution aborts, and why a `Promise.race` timeout can miss a slow body." tag: article - title: New Things You Should Know About HTML Here in Mid 2026 url: https://blog.master.dev/new-things-you-should-know-about-html-here-in-mid-2026/ comment: Chris Coyier tours permission and install elements, customizable selects, scoped custom-element registries, declarative Shadow DOM, HTML-in-canvas, commands, focus groups, and automatic image sizes. Browser support ranges from established to experimental, and the article says so. tag: article - title: WCAG Radar url: https://wcagtoolkit.com/wcag-radar comment: WCAG Radar runs 28 accessibility checks locally in the browser, including on localhost and pages behind a login, without an account or sending page data away. Chrome and Firefox extensions are available; 19 more checks and report export require a licence. tag: library&tools - title: just-secrets url: https://github.com/vercel-labs/just-secrets comment: Vercel Labs wraps native OS credential stores for Node.js 22+ with no dependencies, native add-ons, or install scripts. Storage failures never fall back to files, but the security boundary still depends on the user's OS session and store configuration. tag: library&tools - title: gpu-lexer url: https://gpu-lexer.vercel.app/ comment: Instead of loading language grammars, gpu-lexer uses a 41,321-parameter WebGPU model to infer token classes from context. One 27.4 KB model handles many languages and beats established highlighters in the author's warm benchmark, though accuracy varies sharply by language. tag: ai - title: Sunk Cost url: https://sunkcost.ai comment: "Pick a machine, a model, and daily token volume, then compare buying local hardware with renting the same model by the token. It deliberately answers one narrow question: how long raw inference savings take to repay the box." tag: ai --- My favorite is gpu-lexer—a wonderfully unexpected use of a tiny AI model that delivers real practical value. Sunk Cost also gave me something to think about: whether running LLMs locally can ever pay for itself.