Top Stories

Court Agrees with EFF: Utah’s VPN Law Demands a Technical Impossibility

643 points · eff.org

The day’s runaway top story is a legal win the EFF is understandably eager to publicize: a court has agreed that Utah’s VPN-related law asks for something that simply can’t be built. The crux is a familiar one in the age-verification and content-blocking fights — lawmakers write rules that assume a website can reliably know who’s on the other end of a connection or where they really are, and engineers keep pointing out that VPNs and basic network architecture make that assumption false.

HN has deep, hard-won opinions about the collision between well-meaning internet regulation and how packets actually move, so this one hits a nerve. The thread is the usual mix of people cheering a rare instance of a court listening to technologists and others warning that lawmakers will just try again with slightly different wording.

Apple Pass Designer

438 points · developer.apple.com

Apple has quietly shipped a web-based Pass Designer, a tool for building Apple Wallet passes — boarding passes, loyalty cards, event tickets, coupons — without hand-rolling the JSON and asset bundles that the PassKit format has always required. For anyone who’s fought with the arcane pass.json structure and signing dance, a visual designer that spits out a valid pass is a genuine quality-of-life upgrade.

It’s the kind of developer-tooling release that lands well on HN because it removes a specific, well-remembered papercut. Expect the comments to compare it to the third-party pass generators people have relied on for years, and to probe how much of the signing and distribution workflow Apple is actually smoothing over.

FLUX 3 Image

353 points · bfl.ai

Black Forest Labs is back with FLUX 3, the latest generation of its influential open-weight-adjacent image models. FLUX has been the go-to base model for a huge swath of the open image-generation community, so a new major version is a tentpole event — people want to know how it handles text rendering, prompt adherence, and the fine detail that earlier versions sometimes smeared.

The HN crowd treats each FLUX release as a benchmark for where the open end of image generation stands relative to the closed frontier labs. The thread is already weighing licensing terms, VRAM requirements, and side-by-side comparisons against the previous generation and competing models.

From the Creator of Redis: Run an LLM Locally with ds4

252 points · dwarfstar.sh

Salvatore Sanfilippo — antirez, the creator of Redis — has a new project called ds4 for running LLMs locally, under the Dwarf Star banner. Antirez has a near-mythic reputation in this community for writing lean, legible C and for strong opinions about simplicity, so anything he ships gets read closely regardless of how crowded the space already is.

Local LLM runners are hardly a new category, but the pitch here is clearly about the antirez touch: small, hackable, and built to be understood rather than just used. The comments are part homage to Redis and part genuine curiosity about what a systems programmer of his pedigree thinks the local-inference stack should actually look like.

Sites in ChatGPT

281 points · chatgpt.com

OpenAI is pushing further past the chat box with Sites, a feature that lets ChatGPT generate and host web pages directly. It’s another step in the platform-ward drift of the big AI labs — rather than handing you code to deploy somewhere, the assistant becomes the place the artifact lives, blurring the line between a model and a hosting product.

For HN this raises the strategic questions it loves to chew on: what it means for the open web when a single company owns both the generator and the host, how this stacks up against the wave of “vibe-coding” app builders, and whether developers want their output locked inside someone else’s walled garden. The business subtext — OpenAI expanding surface area into territory that overlaps with Vercel, Netlify, and the no-code crowd — is impossible to miss.

Show HN: Giving Opus 5.5 a Simulated Paint Canvas

283 points · stillwet.art

A Show HN that captured the community’s imagination: hand a frontier model a simulated paint canvas — wet pigment, blending, brush physics — and see what it does when it has to reason about a medium rather than just emit pixels. It’s the kind of playful, probing experiment that reveals something about how these models plan and self-correct when the output isn’t a one-shot generation.

HN loves projects that make a model’s “thinking” legible, and a painting canvas is a particularly vivid way to do it. The thread is full of people sharing their own results and debating whether the model is genuinely reasoning about color and composition or just producing a convincing pastiche.

AI Finally Beat the Best Stratego Player in History — On a Budget

221 points · arstechnica.com

Stratego has long been a holdout for game-playing AI because so much of the board is hidden — unlike chess or Go, you can’t see your opponent’s pieces, so the search space is tangled up with deep uncertainty and bluffing. Ars Technica reports that an AI has now beaten the best human Stratego player in history, and notably did so without the industrial compute budgets that marked earlier game-AI milestones.

That “on a budget” detail is what makes this more than a trophy. HN reads imperfect-information breakthroughs as a better proxy for messy real-world decision-making than perfect-information games ever were, and the comments are digging into the techniques that let a lean system handle hidden state and deception.

Greg Kroah-Hartman — Security in the LLM Age [video]

246 points · youtube.com

Greg Kroah-Hartman, one of the most prominent Linux kernel maintainers, gives a talk on what LLMs mean for security — both the new attack surface they create and what they change for the people maintaining the world’s most scrutinized codebase. When someone who signs off on an enormous volume of kernel patches weighs in on AI-assisted code, the community listens.

The discussion threads the needle HN cares about most: LLMs as a tool for finding and fixing vulnerabilities versus LLMs as a firehose of plausible-looking but subtly broken patches that maintainers now have to triage. Kroah-Hartman’s grounded, maintainer’s-eye view is a useful counterweight to the louder hype in both directions.

The Forgetful CPU (Linux on M4)

222 points · yuka.dev

A deep technical dive into running Linux on Apple’s M4 silicon and the memory-model quirks that make it interesting — the “forgetful CPU” framing points at the subtle reordering and consistency behavior that catches people porting low-level code to Apple’s ARM cores. This is exactly the kind of from-the-trenches systems writeup that HN rewards.

The Asahi-style effort to make Apple Silicon a first-class Linux target has been a years-long saga, and posts that illuminate a specific hardware gotcha — rather than just “it boots now” — are where the real engineering conversation happens. The comments are full of people comparing notes on memory ordering across architectures.