Top Stories

Stacked PRs are now live on GitHub

668 points · github.blog

GitHub has finally shipped native support for stacked pull requests, now in public preview. For years developers have leaned on third-party tools like Graphite and git-branchless to break large changes into a chain of small, dependent PRs that can be reviewed and merged in sequence. Baking this into GitHub itself removes a lot of the friction — no more rebasing headaches or broken base branches when an earlier PR merges.

The HN crowd is enthusiastic but skeptical about the details, since a stacking workflow lives or dies on how gracefully it handles rebases, merges, and force-pushes. It’s a meaningful nod to how large teams actually ship code, and a signal that GitHub is defending its turf against the growing ecosystem of purpose-built stacking tools.

Gemini Robotics 2 brings whole-body intelligence to robots

568 points · deepmind.google

Google DeepMind’s latest robotics model extends its vision-language-action approach to “whole-body” control — coordinating locomotion and manipulation together rather than treating a robot’s arms and legs as separate problems. The pitch is a single foundation model that can reason about a task and then execute the full-body motion to accomplish it.

This is one of the most-watched frontiers in AI right now: everyone from DeepMind to Figure to Tesla is racing to make general-purpose robots actually useful. Commenters are debating how much of the flashy demo generalizes to messy real-world environments versus carefully staged setups, a recurring tension in robotics announcements.

DeepSeek-V4-Flash update

331 points · api-docs.deepseek.com

DeepSeek has rolled out V4-Flash, a fast, cheap variant of its V4 model line, and the independent benchmarks are already flowing in. Early analysis from Artificial Analysis puts it in a strong price-performance spot, which is exactly the wedge that has made DeepSeek a thorn in the side of larger US labs.

HN readers care because DeepSeek keeps demonstrating that competitive models can be trained and served at a fraction of the assumed cost, pressuring incumbents on pricing. There’s a companion “Show HN” on the front page showing that distilling DeepSeek into an open model doesn’t carry over its censorship behavior — a reminder of how much geopolitics is baked into these releases.

GCC steering committee announces AI policy

304 points · lwn.net

The GCC steering committee has published a policy on AI-generated contributions to the compiler, joining a growing list of major open-source projects trying to get ahead of the flood of machine-written patches. The core questions are about copyright provenance and the Developer Certificate of Origin: can a contributor legitimately sign off on code a model produced?

This is a live wire in the open-source world, where maintainers are already drowning in low-quality “AI slop” pull requests. GCC’s stance matters because it’s one of the most consequential codebases on the planet, and other projects will look to how it threads the needle between welcoming productivity gains and protecting the license integrity of the code.

The AI Aesthetic

315 points · blog.jim-nielsen.com

A sharp essay on the emerging visual and interaction “look” of AI-generated software — the sameness that creeps in when everyone builds on the same models, prompts, and default component libraries. Nielsen argues that the tools nudge us toward a homogenized aesthetic, and that resisting it takes deliberate taste and effort.

It struck a nerve with designers and developers who’ve noticed how many recent products feel interchangeable. The comments turn into a broader debate about whether AI tooling flattens creativity or simply lowers the floor while leaving the ceiling untouched — a theme that keeps resurfacing as generative tools spread through the design world.

Read this before you buy that TV streaming stick

725 points · krebsonsecurity.com

Brian Krebs digs into the security nightmare lurking inside cheap, off-brand Android streaming devices — the kind sold for a few dollars on big marketplaces. Many ship with pre-installed malware or backdoors baked into the firmware, quietly enrolling your home network into botnets and ad-fraud schemes before you’ve even finished setup.

It’s a classic Krebs investigation and it resonated hard with the HN audience, which has long warned about the supply-chain risks of ultra-cheap connected hardware. The practical takeaway — be wary of no-name streaming boxes and stick to reputable brands — is the kind of concrete advice readers can pass along to less technical family members.

I flagged two research papers for fake authors and both were accepted as orals

217 points · geospatialml.com

A reviewer recounts flagging two conference submissions for apparently fabricated authors and AI-generated content — and watching both get accepted as oral presentations anyway. It’s a damning anecdote about the state of peer review as the volume of machine-assisted submissions explodes.

The story landed because it crystallizes a fear many academics share: that review pipelines are buckling under the weight of AI-generated “slop,” and that acceptance decisions are increasingly disconnected from rigor. Commenters swap their own war stories about broken review processes and debate what incentives or tooling could realistically fix it.

The Economic Benefit of Refactoring

246 points · martinfowler.com

Part of Martin Fowler’s ongoing series on generative AI and software, this piece makes the case that refactoring pays for itself by keeping code cheap to change — and that the argument becomes even more important in an era where AI can generate lots of code fast. Well-structured code is easier for both humans and models to work in.

Fowler is a perennial HN favorite, and the discussion turns into the familiar-but-worthwhile debate over how to justify refactoring to skeptical managers who see it as time not spent shipping features. The AI angle adds a fresh twist: if models generate the code, does design discipline matter more or less?

JEP 401: Value Objects (Preview) merged to OpenJDK master

131 points · github.com

A milestone for Project Valhalla: JEP 401’s value objects have been merged into the OpenJDK master branch as a preview feature. Value objects let developers define classes whose identity doesn’t matter — enabling the JVM to flatten and optimize them for big performance and memory wins without exposing pointer identity.

Java watchers have been waiting the better part of a decade for Valhalla to land, so any concrete merge is a cause for cautious celebration. The comments dig into what this means for real-world performance and how it interacts with existing code, along with the usual reflections on just how long ambitious JVM projects take to ship.