Top Stories

Steam Machine launches today

1616 points · store.steampowered.com

Valve’s long-rumored living-room PC is finally real and shipping. The new Steam Machine lands as the runaway top story of the day, and the HN crowd is treating it as a genuine moment — Valve has tried the console-PC hybrid before and stumbled, so a credible, polished second attempt is a big deal. The interest isn’t just nostalgia: with SteamOS now battle-tested on the Steam Deck, a fixed hardware target gives developers something the original 2015 Steam Machines never had, and revives the perennial “year of Linux gaming” debate.

GLM-5.2 – How to Run Locally

415 points · unsloth.ai

Another strong open-weight release lands with a practical local-inference guide attached, and that combination is exactly what HN loves. GLM-5.2 continues the trend of frontier-adjacent Chinese models shipping with permissive weights, and Unsloth’s quantization-and-run instructions mean people can actually try it on their own hardware rather than just reading benchmarks. The thread is the usual mix of VRAM math, quant tradeoffs, and comparisons against the latest closed models.

VibeThinker: 3B param model that beats Opus 4.5 on reasoning

213 points · arxiv.org

A 3-billion-parameter model claiming to outscore Opus 4.5 on reasoning benchmarks is the kind of result that makes everyone squint at the eval methodology — and that’s half the fun of the thread. The paper credits a novel SFT-plus-GRPO training recipe, feeding the ongoing argument that clever post-training, not raw scale, is where the next gains live. If the numbers hold up outside the benchmark, a model this small running reasoning workloads locally would be a meaningful shift.

Canada plans ‘nuclear renaissance’ with up to 10 reactors by 2040

494 points · cbc.ca

Canada’s federal government laid out a strategy to build as many as ten new reactors by 2040, and the timing dovetails with the broader tech-industry conversation about where the electricity for AI data centers is going to come from. HN’s energy contingent is engaged on the SMR-versus-large-reactor question and on whether the timelines are realistic given how nuclear projects have historically slipped. It’s a reminder that the AI buildout is increasingly an energy-policy story.

Moebius: 0.2B image inpainting model with 10B-level performance

297 points · hustvl.github.io

A 0.2B-parameter inpainting model punching at the level of models 50x its size is another data point in the “small and efficient” wave dominating today’s front page. Inpainting is a workhorse task — object removal, photo cleanup, generative fill — so a tiny model that matches heavyweight quality is immediately practical for on-device and edge use. The demos drew the usual scrutiny over cherry-picking, but the efficiency claim is what has people talking.

Show HN: Oak – Git alternative designed for agents

190 points · oak.space

As coding agents become first-class users of version control, someone has finally asked whether Git’s human-centric model is the right fit — and built an alternative around how agents actually work. Oak sparked a lively debate about whether the friction agents hit with Git (merge conflicts, staging, history rewriting) warrants a new tool or just better tooling on top of Git. Either way, “version control designed for agents” captures where a lot of developer-tools thinking is heading.

The new HTTP QUERY method explained

139 points · kreya.app

HTTP is getting a new QUERY method, and this explainer walks through why it exists: a safe, idempotent, body-carrying request that finally solves the awkward “GET with a body” problem developers have hacked around for years. Anyone who has stuffed a complex search filter into a giant URL or quietly switched to POST for read-only queries will recognize the pain it addresses. The thread digs into caching semantics and how long real-world framework and proxy support will take.

In praise of memcached

174 points · jchri.st

A love letter to memcached struck a chord with engineers tired of reaching for heavier, more featureful caches by default. The argument is that memcached’s deliberate simplicity — do one thing, do it fast, stay boring — is a feature, not a limitation, and that plenty of systems pull in Redis for capabilities they never use. It’s the kind of “old, unglamorous, still excellent” infrastructure post HN reliably rallies around.

OpenAI DayBreak – GPT-5.5-Cyber

117 points · openai.com

OpenAI announced a cybersecurity-focused variant of its latest model, pitched at defensive security work. The framing — a frontier model specialized for “securing the world” — landed in a skeptical thread weighing genuine blue-team utility against the obvious dual-use concern that the same capabilities help attackers. It’s part of a growing push to position AI as a security tool rather than just a security risk, and HN isn’t taking the marketing at face value.

Flock-powered police chiefs stalking women shows why warrants are needed

527 points · ipvm.com

An investigation into police officials misusing Flock’s automated license-plate-reader network to track women crystallizes HN’s long-running unease about pervasive surveillance infrastructure. The reporting is a concrete example of the “build the dragnet, abuse follows” warning privacy advocates have made for years, and it has reignited the debate over warrant requirements and oversight for ALPR systems. Expect strong feelings about surveillance capitalism and civil liberties in the comments.