My favorite is the whole Jev thing. It feels like one of the biggest steps forward in AI: some tasks can be solved much faster and more cheaply. I hope it leads to some great new software in the coming months.
News
Safari 27’s release notes are the practical compatibility checklist: new web-platform features, changed behaviour, and fixes in one place. Worth scanning before declaring a browser quirk solved or a new API safely universal.
Voting is open for this year’s 13 kB JavaScript game entries. The constraint remains a lovely antidote to modern toolchains: a complete game, art, audio, and code must all fit in a tiny downloadable package.
Article
A visual map of functional-programming terms, with JavaScript examples to keep words such as Alternative from becoming a private dialect. Useful when a library’s documentation assumes you already speak category theory.
Linear cut runner time per test roughly in half while its suite nearly quadrupled. The useful details are mundane: faster runners, smaller checkouts, fewer installs, fewer cache rituals, and getting non-gating work off the critical path.
This paper gives Skia 2D rasterization a formal Lean semantics, then verifies optimizations for inefficient instruction sequences found in Chrome. Its optimizer reports an 18.7% speedup across 99 programs with 32 microseconds of optimization work.
A compact museum of the star hack, `hasLayout`, CSS expressions, `cursor: hand`, custom IE scrollbars, and HTML Components. Great nostalgia, and a reminder that today’s browser compatibility complaints are comparatively civilized.
Eric Eggert’s critique of the new WCAG 3 draft is precise: core conformance is clearer, but optional Supplemental Requirements and Assertions risk making a difficult standard longer without making organisations do more accessibility work.
Library&tools
This Oxc release updates React lint rules for React 19.3, fixes a stack of diagnostics, and removes small allocations from hot rules. Oxfmt also gains prefix alignment and several comment-handling fixes.
Lovable is replacing Vite instances in its preview fleet with OJ, a Rust binary built on Rolldown and Oxc. It reads Vite config and bridges real plugins, but targets the very different economics of millions of disposable sandboxes.
Fallow is static codebase intelligence for TypeScript and JavaScript: dead code, duplication, complexity, architecture drift, and design-system drift from one module graph. It starts with `npx fallow`, then adds runtime evidence only when deletion confidence needs it.
UI Skills is a searchable catalogue of design-engineering skills for agents, with a CLI and MCP entry point. It covers focused work such as motion, accessibility, layout, visual hierarchy, and interface reviews rather than one vague “make it pretty” prompt.
Ai
fx is an experimental, open-source Zig coding-agent harness that fits in a 6.17 MiB binary. It is model- and provider-agnostic, can run in WebAssembly, and deliberately feels more like a Unix tool than an IDE.
TypeSafe’s Jev is a closed early-access model for structured decisions: it selects typed options instead of generating explanatory text. The pitch is fast, calibrated routing and classification, while accepting that it cannot produce ordinary free-form answers.
SemIf is an independent open-model experiment in the same decision-first shape: read option logits directly, skip token generation, and reuse shared state across questions. Its authors are clear that it reproduces the interface pattern, not Jev’s undisclosed model.
Browser Use pairs Jev’s operation and target selection with a small text model only for typing. Its Google Flights demo finished in 7.1 seconds, though the repository sensibly labels its evidence as a few runs of one task, not a general benchmark.