Top Stories

GLM 5.2 beats Claude in our benchmarks

879 points · semgrep.dev

Semgrep ran its internal cyber-security benchmark suite against Zhipu’s newly released GLM 5.2 and found it edging out Claude on their offensive- and defensive-security tasks. The headline is catnip for HN: an open-weight Chinese model topping a frontier US lab on a real, applied evaluation rather than a generic leaderboard. The comments dig into the usual caveats — benchmark contamination, task selection, and whether “beats Claude” holds outside this narrow domain — but the broader signal is hard to ignore. The gap between open and closed models keeps shrinking, and teams that build security tooling are increasingly willing to run whichever model wins on their own evals.

Age verification is just a precursor to automated attribution of speech

546 points · nonogra.ph

This essay argues that the current wave of age-verification mandates isn’t really about protecting kids — it’s the on-ramp to tying every utterance online back to a verified identity. Once the infrastructure exists to prove “you are over 18” on every site, the same plumbing can prove “you are this specific person,” and anonymous speech quietly stops being an option. The piece resonated with the HN crowd’s long-standing privacy anxieties, landing right alongside the KIDS Act discussion below. Commenters debate whether the slippery-slope framing is alarmist or simply describing the obvious endgame of identity-gated internet access.

HackerRank open sourced its ATS. My resume scored 90/100. Oh wait 74. No – 88

510 points · danunparsed.com

HackerRank open-sourced the applicant-tracking and resume-scoring system it sells to employers, and the author promptly fed their own resume through it — getting wildly different scores (90, then 74, then 88) across runs on the same document. It’s a brutal, concrete illustration of how much weight hiring pipelines place on non-deterministic, often LLM-backed scoring that can’t even agree with itself. For a community that spends a lot of energy resenting automated resume screens, the schadenfreude is thick. The deeper point lands too: if the gatekeeping tool is this noisy, plenty of qualified candidates are being filtered out by what amounts to a coin flip.

The KIDS Act would require age checks to get online

504 points · eff.org

The EFF breaks down the proposed KIDS Act, which would push mandatory age verification across broad swaths of the internet. Their argument is the familiar civil-liberties case: age gates force everyone — not just minors — to hand over identifying data, create honeypots of sensitive information, and chill anonymous participation. Paired with the attribution-of-speech essay above, this was the policy story of the day on HN. The thread is a mix of “think of the kids is always the wedge” cynicism and genuine debate over whether any privacy-preserving age-verification scheme can actually work at scale.

I used Claude Code to get a second opinion on my MRI

447 points · antoine.fi

The author walks through pointing Claude Code (running Opus) at their own MRI imaging and report to get a layperson-readable second opinion — translating radiology jargon, flagging what to ask their doctor, and cross-referencing findings. It’s equal parts impressive and unnerving, and the HN thread reflects that tension: stories of the model catching context a rushed appointment glossed over, against sober warnings about hallucinated findings and the danger of self-diagnosis. It captures a real shift in how people are reaching for general-purpose AI tools in moments that used to require a specialist — and the very real risks of doing so.

Professor denounces mass AI fraud on an exam at Brown

421 points · english.elpais.com

A Brown University professor went public after discovering widespread AI-assisted cheating on an exam, reigniting the now-perennial debate over what assessment even means in the LLM era. The HN discussion splits along predictable lines: some argue the answer is a return to in-person, closed-book exams; others insist that any task a chatbot can complete was never worth grading in the first place. Underneath the noise is a real institutional crisis — universities built their integrity systems for a world where producing fluent text was itself proof of effort, and that assumption no longer holds.

Librepods: AirPods liberated

404 points · github.com

Librepods is an open-source project that reverse-engineers Apple’s AirPods protocol to bring features like ear-detection, battery status, and spatial-audio toggles to non-Apple platforms — Android and Linux included. It’s a classic HN favorite: reclaiming hardware you own from a walled garden, documented in the open. The thread is full of appreciation for the reverse-engineering work plus the usual wariness about how long Apple will tolerate it before a firmware update breaks things. For anyone who’s resented their AirPods going dumb the moment they leave the Apple ecosystem, it’s a welcome bit of liberation.

Historical memory prices 1960-2026

321 points · dam.stanford.edu

A meticulously maintained dataset and visualization tracking the price of computer memory from 1960 to today, charting the staggering multi-order-of-magnitude collapse in cost per byte. It’s the kind of long-horizon data that HN loves to pore over, and the comments turn it into a tour of computing history — punch cards, core memory, the DRAM cartels, and the recent plateau. There’s also a timely subtext: with AI training and inference devouring high-bandwidth memory, several commenters note that the decades-long downward trend has gotten bumpy, and HBM in particular is bucking it hard.

A way to exclude sensitive files issue still open for OpenAI Codex

210 points · github.com

This GitHub issue — still open — asks OpenAI’s Codex agent for a reliable way to exclude sensitive files (think .env, secrets, private keys) from being read and potentially exfiltrated to the model. That such a basic guardrail is unresolved set off the HN security crowd. As coding agents get more autonomous and routinely crawl entire repositories, the lack of a robust, default-on exclusion mechanism feels like a glaring gap. The thread is a useful reminder that the convenience of agentic dev tools is racing ahead of the boring-but-critical safety controls underneath them.