Top Stories

Your ePub Is Fine — Kobo Disagrees, Blame Adobe

650 points · andreklein.net

A deep dive into why perfectly valid ePub files get rejected by Kobo readers, with the trail leading back to Adobe’s aging DRM and validation stack. It’s the kind of forensic debugging story HN loves: a developer methodically proving their file is correct and the platform is wrong. The thread also reignited the perennial gripe about ebook DRM holding back an otherwise open, simple format.

Firewood Splitting Simulator

842 points · screen.toys

The runaway crowd favorite of the day — a browser toy that does exactly what it says, with oddly satisfying physics. It topped the front page on pure delight, a reminder that a tight, polished interactive demo can outscore every serious tech story on HN. Expect a wave of “I lost an hour to this” comments.

Show HN: Kage – Shadow Any Website to a Single Binary for Offline Viewing

576 points · github.com

Kage crawls a site and packages it into one self-contained binary you can run offline. The single-binary distribution angle is what resonated — no dependencies, no server setup, just a file you can archive or ship. Commenters discussed use cases from documentation snapshots to preserving sites at risk of going dark, plus the usual debate about respecting robots.txt and rate limits.

Curl Will Not Accept Vulnerability Reports During July 2026

452 points · daniel.haxx.se

Daniel Stenberg announced curl is pausing its security report intake for all of July, framing it as a much-needed break for a tiny maintainer team. The subtext everyone seized on: the flood of low-quality, AI-generated “vulnerability” reports that have been drowning open-source maintainers. It crystallizes a growing tension about who bears the cost when LLMs make it trivial to file plausible-looking but worthless security claims.

Windows 11 Users Are Tired of Microsoft Account Requirements Creeping into Everything

382 points · windowscentral.com

Microsoft keeps closing the loopholes that let people set up Windows 11 with a local account, and users are increasingly fed up. The HN reaction was a mix of resignation and migration talk — Linux distros, debloat scripts, and the broader question of how much control you really have over a machine you bought. A reliable flashpoint for the “your computer isn’t yours anymore” debate.

Rio de Janeiro’s “Homegrown” LLM Appears to Be a Merge of an Existing Model

359 points · github.com

A GitHub issue lays out evidence that a much-hyped “homegrown” Brazilian LLM is largely a merge of existing open-weight models rather than something trained from scratch. The story taps into a recurring AI theme: governments and orgs claiming sovereign model achievements that don’t hold up under scrutiny. Commenters dug into the technical fingerprints — tokenizer artifacts, weight similarities — that give merges away.

Formal Methods and the Future of Programming

280 points · blog.janestreet.com

Jane Street published an index of its work on formal methods, making the case that mathematically proving software correct is becoming more practical for real systems. Given Jane Street’s reputation and OCaml pedigree, the post drew serious discussion about where formal verification actually pays off versus where it’s still too costly — and whether AI-assisted proof tooling could finally widen its reach.

Apple Foundation Models

183 points · platform.claude.com

Documentation surfaced for using Apple’s on-device Foundation Models, including bindings within Claude’s SDK ecosystem. The interest centers on the practical reality of local, privacy-preserving inference on Apple silicon and how it slots into existing developer toolchains. It’s another data point in the shift toward running capable models directly on user hardware rather than round-tripping to the cloud.

The Only Scalable Delete in Postgres Is DROP TABLE

172 points · planetscale.com

PlanetScale argues that large-scale row deletion in Postgres is fundamentally painful — bloat, vacuum pressure, lock contention — and that partitioning so you can simply drop whole tables is the only approach that truly scales. It’s a meaty database-engineering read, and the thread filled with war stories about TTL data, retention policies, and the dark art of keeping Postgres happy at volume.