Top Stories

2D Vehicles

496 points · patkerr.co.uk

Sometimes the front page just wants to have fun, and this interactive playground of physics-driven 2D vehicles is the kind of thing that pins HN for a day. It’s a browser toy where you can tinker with wheels, suspension, and terrain and watch the little contraptions scramble around — the sort of hands-on demo that reminds engineers why they got into this in the first place. The crowd loves it because it’s immediate, polished, and endlessly fiddleable, no sign-up or install required.

The Lightbulb Computer

440 points · lightbulbcomputer.com

A computer built out of incandescent lightbulbs is exactly the kind of beautifully impractical hardware project HN can’t resist. The piece walks through using the nonlinear behavior of lightbulb filaments as logic elements — turning a humble household object into computational primitives. It lands squarely in the “because I can” tradition of relay computers and marble machines, and the comments are full of people marveling at the ingenuity while debating whether the filaments really behave the way the author claims.

A city-building game in which the city would prefer you didn’t

361 points · housing.over.pizza

An adversarial city-builder with a deadpan premise — the city actively resists your attempts to develop it — is both a clever game design hook and a sly bit of commentary on zoning and NIMBYism. It plays like a puzzle where every improvement triggers new obstacles, and the housing-policy subtext is not subtle. HN’s developers and armchair urbanists are having a field day both with the mechanics and with the real-world politics it pokes at.

Build your own decision model

324 points · nishtahir.com

A practical walkthrough of building a personal decision-making framework resonated hard with the engineering crowd, who tend to want to systematize even their life choices. Rather than handing over a rigid template, the author shows how to structure the factors, weights, and tradeoffs that go into a hard call so you can reason about it explicitly. It’s catnip for a community that respects clear thinking, and the comments turn into a swap meet of everyone’s own decision heuristics.

Talorys – A self-hosted personal AI agent on Cloudflare’s free tier

298 points · github.com

This one checks every HN box at once: a self-hostable AI agent you can run entirely on Cloudflare’s free tier, no cloud bill required. In a moment when everyone is trying to own their AI stack rather than rent it from a hyperscaler, a project that gives you a personal agent at zero marginal cost is instantly appealing. The appeal is as much about control and privacy as it is about price — and the near-zero comment count relative to its points suggests people are busy cloning the repo rather than arguing about it.

Mxc: Microsoft Execution Containers version 1.0.0

194 points · blogs.windows.com

Microsoft shipping 1.0 of a policy-driven containment layer specifically for AI agents is a signal of where the industry’s head is at: as agents gain the ability to run code and touch systems, sandboxing them is becoming table stakes. Mxc is pitched as a way to give agents real capabilities while fencing in what they can actually do. HN is dissecting the security model closely — part genuine interest in the containment primitives, part healthy skepticism about trusting any agent with execution rights at all.

Nix wrote half of my debugger

200 points · fzakaria.com

A love letter to Nix from someone who discovered that its reproducible, fully-specified build environments did a surprising amount of the heavy lifting in building a debugger. The post is a concrete case study in how deterministic tooling pays off: when your dependencies and toolchain are pinned exactly, a whole class of “works on my machine” debugging problems simply evaporates. It’s the kind of real-world Nix win that converts skeptics, and the thread is the usual mix of Nix evangelists and people still bruised by its learning curve.

Five months treating bugs like patients and coding agents like a medical team

140 points · cockroachlabs.com

Cockroach Labs ran a five-month experiment organizing their bug-fixing workflow like a hospital — triage, specialists, rounds — with coding agents playing the roles on the medical team. It’s one of the more thoughtful write-ups of actually operationalizing AI agents inside a serious engineering org, going well past the demo stage into process design and metrics. For HN, which is deeply ambivalent about AI-assisted development, it’s a welcome dose of real data on what worked, what didn’t, and where the agents still needed a human attending physician.

Retrofitting language models to operate over bytes

71 points · nature.com

A Nature paper on retrofitting existing language models to work directly over raw bytes rather than tokens is the kind of substantive ML research HN rewards. Byte-level operation sidesteps a pile of tokenizer headaches — vocabulary limits, multilingual quirks, and brittle handling of code and structured data — and doing it as a retrofit rather than a from-scratch retrain is the practical twist. The discussion digs into whether the efficiency tradeoffs are worth it and what this means for the long-running tokenizer-versus-bytes debate.