Top Stories

Noise Infusion Banned from Census Bureau Statistical Products

830 points · desfontain.es

After years of fierce debate over the 2020 Census’s use of differential privacy, a new directive bans the deliberate injection of statistical noise into published Census data. The author — a differential-privacy researcher — walks through what the policy actually says and why it’s a watershed moment for the field. HN is split: privacy folks see a hard-won protection being gutted, while data users who fought the noisy 2020 tables are quietly celebrating.

Every Frame Perfect

733 points · tonsky.me

Nikita Prokopov (Tonsky) digs into what it really takes to render UI without dropped frames or visual tearing, from vsync and frame pacing to the subtle ways modern toolkits get it wrong. It’s the kind of meticulous, screenshot-heavy deep dive HN loves, and the comments turned into a master class on display timing, compositors, and why “60fps” is harder than it sounds.

Amazon CEO’s Talks with U.S. Officials Triggered Crackdown on Anthropic Models

698 points · wsj.com

The WSJ reports that conversations between Amazon’s CEO and U.S. officials helped set off the government directive now restricting access to certain Anthropic models. It’s a striking look at how big-tech lobbying and frontier-AI regulation are colliding behind closed doors, and the thread is full of speculation about competitive motives, national-security framing, and what it means for everyone building on these APIs.

GLM 5.2 Is Out

611 points · twitter.com

Zhipu’s GLM 5.2 is here, and the announcement claims meaningful gains on coding and reasoning benchmarks that put it in striking distance of the top Western frontier models. With open-weight Chinese models advancing this fast, HN’s debate centers on benchmark trustworthiness, real-world coding performance, and whether the open-weight ecosystem is now genuinely competitive with closed labs.

Treating Pancreatic Tumours May Have Revealed Cancer’s Master Switch

378 points · economist.com

Researchers working on one of the deadliest cancers stumbled onto a regulatory mechanism that may act as a broader “master switch” across tumour types. The Economist lays out the science and the appropriate caveats, and the HN crowd — including a few biologists — weighs in on how promising preclinical signals like this often look before the hard road of trials.

AI Coding at Home Without Going Broke

300 points · stephen.bochinski.dev

A practical guide to getting real value from AI coding assistants without burning through expensive API credits — covering local models, smart context management, and when to reach for the pricey frontier APIs versus a good-enough local setup. It resonated with developers tired of surprise bills, and the comments are a goldmine of cost-control tactics and self-hosting tradeoffs.

A Low-Carbon Computing Platform from Your Retired Phones

298 points · research.google

Google Research describes turning fleets of retired smartphones into a low-carbon distributed computing platform, reusing hardware that would otherwise be e-waste. It’s a clever sustainability-meets-systems story, and HN dug into the practicality — battery degradation, security, orchestration — alongside genuine enthusiasm for keeping capable silicon out of landfills.

RTX 5080 + RTX 3090 Setup: 80 Tok/s on Qwen 3.6 27B Q8

254 points · imil.net

A detailed homelab writeup on mixing a current-gen RTX 5080 with an older 3090 to run Qwen 3.6 27B at Q8 quantization, hitting 80 tokens per second. It’s catnip for the local-LLM crowd, and the thread dives into mismatched-GPU pitfalls, VRAM splitting, quantization quality, and whether mixed setups are worth the hassle versus a single big card.

Codex for Open Source

237 points · openai.com

OpenAI is opening up Codex access for open-source projects, pitching AI-assisted development as a boost for maintainers stretched thin. Reactions range from maintainers eager for help triaging issues and PRs to skepticism about AI-generated contributions adding review burden — a live debate about whether these tools genuinely lighten the OSS maintenance load or just shift it.