Top Stories

Linux gaming is faster because Windows APIs are becoming Linux kernel features

763 points · xda-developers.com

The Year of the Linux Desktop arrives through the side door of gaming. XDA digs into how Proton, Wine, and increasingly the kernel itself are absorbing Windows API behaviors — to the point where many titles now run faster on Linux than on the OS they were written for. Valve’s investment in the Steam Deck is the proximate cause, but the second-order effects on the wider kernel are what has HN buzzing.

The community discussion centers on whether this represents a genuine inflection point or just a niche win for a subset of well-optimized titles. Either way, the trajectory is unmistakable: Linux is no longer reverse-engineering Windows so much as it’s quietly outperforming it on its own games.

Show HN: Needle — Distilling Gemini Tool Calling into a 26M Model

678 points · github.com

Cactus Compute released Needle, a 26-million-parameter model distilled specifically for tool calling that they claim matches Gemini’s performance on the task. At that size, it runs comfortably on-device — which is the whole point. As agentic workflows proliferate, the bottleneck isn’t raw reasoning but cheap, fast, reliable function calls, and a sub-50M model that does that one job well is a meaningful unlock.

HN’s response is a mix of excitement about the on-device implications and pointed questions about eval methodology and whether tool calling is really separable from broader reasoning. Expect a wave of clones if the benchmarks hold up.

Setting up a free *.city.state.us locality domain (2025)

569 points · fredchan.org

A delightful piece of internet archaeology turned how-to: it turns out US municipalities can still request free city.state.us domains under a long-standing IANA delegation. Fred Chan walks through the process, including which states are responsive and which have effectively gone dormant. It’s the kind of forgotten infrastructure HN loves to rediscover.

The thread is full of stories from civic technologists who’ve used these domains for community projects, plus laments about commercial gTLDs eating the namespace that this older system was meant to protect.

Twin brothers wipe 96 government databases minutes after being fired

447 points · arstechnica.com

A cautionary tale that reads like a security policy in-joke made flesh: two sysadmin twins, fired the same day, used still-live credentials to drop tables across 96 government databases before anyone could revoke access. Ars covers the prosecution and what it says about offboarding hygiene in public-sector IT.

For HN readers this is mostly a Rorschach test — half the comments are “this is why you rotate credentials at termination,” the other half are “this is why you don’t fire people via Slack with their VPN still live.”

A History of IDEs at Google

380 points · laurent.le-brun.eu

A long-form insider history of how Google’s developer environment evolved from Emacs/Vim on workstations through Mondrian, Critique, Cider, and the cloud-IDE era. Laurent Le Brun, who worked on several of these systems, traces the engineering trade-offs and political battles that shaped each generation.

It’s catnip for the dev-tools crowd, and a useful corrective to the idea that Big Tech’s internal tooling is uniformly superior — much of the story is about how hard it is to retire beloved-but-creaky systems.

Princeton mandates proctoring for in-person exams, upending 133-year precedent

333 points · dailyprincetonian.com

Princeton’s faculty voted to end the university’s 133-year-old unproctored honor-code tradition for in-person exams. The trigger, predictably, is AI: the existing system assumed students could be trusted not to consult outside resources, but in 2026 “outside resources” includes a phone in their pocket that can solve almost any problem set.

HN’s discussion is wide-ranging — the death of honor systems, what assessment even means now, and whether the right response is more proctoring or fundamentally different exam formats. The answer is probably “both,” but neither is cheap.

The Emacsification of Software

315 points · sockpuppet.org

Thomas Ptacek argues that an increasing share of modern developer tools — VS Code, Cursor, Zed, and the agentic IDEs — are recapitulating Emacs’s core insight: a small, scriptable kernel surrounded by a vast ecosystem of user-extensible modes. The piece is part celebration, part warning that we’re rebuilding Emacs poorly.

The comment section is unsurprisingly spicy. Emacs partisans nod along; the Cursor crowd pushes back; nobody comes out looking neutral.

Claude for Small Business

285 points · anthropic.com

Anthropic launches a packaged offering aimed squarely at SMBs — bundled seats, simplified billing, and templated workflows for the kinds of tasks small teams actually do (drafting proposals, summarizing customer threads, running through compliance checklists). It’s an explicit move down-market from the enterprise pricing that has dominated Anthropic’s revenue mix so far.

The strategic read: the SMB segment is where ChatGPT Team has been growing fastest, and Anthropic doesn’t want to cede the long tail. Expect Microsoft and Google to respond within weeks.

Reverting the incremental GC in Python 3.14 and 3.15

239 points · discuss.python.org

The Python core team is reverting the incremental garbage collector that landed in 3.14 after benchmarks and bug reports showed regressions in real-world workloads, particularly for long-running web services. It’s a rare public walk-back of a flagship feature and the proposal lays out exactly what went wrong and what’s planned for 3.16.

HN’s take is mostly admiring: “this is what good engineering looks like” energy, with side conversations about why GC tuning is so brittle and whether Python should bite the bullet on a generational rewrite.

The US is winning the AI race where it matters most: commercialization

205 points · avkcode.github.io

A counter-narrative to the “China is catching up” discourse: the argument is that benchmark-leading models matter less than which country actually monetizes AI in production. By that lens, US dominance is widening, not shrinking, with most of the world’s AI-driven revenue running through US-owned platforms regardless of where the underlying weights were trained.

It’s a contested thesis and the comments don’t pull punches — readers debate whether commercialization advantage is durable or whether open-weights releases are quietly eroding it.