Top Stories

S&P 500 Rejects SpaceX, Also Blocking OpenAI and Anthropic

543 points · arstechnica.com

S&P Dow Jones Indices declined to fast-track SpaceX into the S&P 500 and reaffirmed the rules that keep unprofitable companies out — a decision that also sidelines OpenAI and Anthropic. For all the hype around the most valuable private tech companies of the era, profitability still gates entry to the index that anchors trillions in passive investment.

The HN crowd is dissecting what this says about the gap between paper valuations and actual earnings. With index inclusion driving enormous automatic inflows, the snub is a reminder that the most talked-about AI and space firms remain, by the index committee’s standard, financially unproven.

Mouseless – Keyboard-Driven Control of macOS, Linux and Windows

536 points · mouseless.click

Mouseless is a cross-platform tool that lets you drive your entire desktop from the keyboard — clicking, navigating, and interacting without ever reaching for the trackpad. It’s the kind of power-user utility that HN reliably rallies around, and the cross-OS support is what set this one apart.

The appeal is partly ergonomic and partly about flow: keeping your hands on home row eliminates the constant context-switch to the mouse. Commenters are comparing it to Vimium, Homerow, and other keyboard-navigation tools, debating how well it handles apps that weren’t designed with keyboard control in mind.

Gov.uk Has Replaced Stripe with Dutch Provider Adyen

468 points · theregister.com

The UK government’s digital services platform has swapped out Stripe in favor of Amsterdam-based Adyen for payment processing. It’s a notable enterprise win for Adyen and a rare public-sector loss for Stripe, and the timing dovetails with broader European conversations about tech sovereignty.

The discussion is split between the procurement angle — why a government would move a payments stack that presumably worked — and the geopolitics of routing public payments through a European rather than American processor. Either way, displacing an incumbent of Stripe’s stature on a platform this visible is a big deal.

Did Claude Increase Bugs in rsync?

427 points · alexispurslane.github.io

This analysis digs into whether AI-assisted contributions correlated with a rise in bugs in rsync, one of the most battle-tested pieces of Unix plumbing. It’s exactly the kind of empirical, name-a-real-project case study the community wants more of as AI coding assistants spread into critical infrastructure.

The debate is sharp: some read it as evidence that LLM-generated patches introduce subtle regressions into mature codebases, while others push back on the methodology and causation. Coming alongside the day’s “why is HN so anti-AI?” thread, it captures the moment’s central anxiety about trusting machines with code that everything depends on.

pg_durable: Microsoft Open-Sources In-Database Durable Execution

407 points · github.com

Microsoft has open-sourced pg_durable, a Postgres extension that brings durable execution — the workflow pattern popularized by tools like Temporal — directly into the database. Instead of running a separate orchestration service, you get crash-resistant, resumable workflows where the database itself is the source of truth.

Developers are excited about collapsing a whole tier of infrastructure into Postgres, which already does so much heavy lifting in modern stacks. The conversation weighs the appeal of fewer moving parts against the classic question of whether the database should be taking on yet another responsibility.

How LLMs Work

392 points · 0xkato.xyz

A from-the-ground-up explainer of how large language models actually function — tokens, embeddings, attention, and the mechanics of next-token prediction — written for people who want the real mental model rather than a hand-wavy metaphor. These deep-but-accessible primers consistently do well on HN.

Readers are praising the clarity while debating where explanations like this should draw the line between intuition and mathematical rigor. As LLMs move from novelty to daily tool, demand for genuinely good conceptual explanations keeps climbing.

Gemma 4 QAT Models: Optimizing Compression for Mobile and Laptop Efficiency

357 points · blog.google

Google released quantization-aware-trained versions of Gemma 4, squeezing the open models down so they run efficiently on phones and laptops without the usual accuracy hit from post-hoc quantization. QAT bakes the low-precision constraints into training itself, preserving more quality at smaller sizes.

This lands squarely in the local-LLM movement, where running capable models entirely on-device — no API, no data leaving your machine — is the goal. Commenters are comparing footprint and benchmark numbers against other open models and swapping notes on what’s now practical to run locally.

Conventional Commits Encourages Focus on the Wrong Things

318 points · sumnerevans.com

A contrarian take arguing that the popular Conventional Commits spec — the feat:/fix:/chore: prefix convention — pushes developers to obsess over commit formatting instead of writing genuinely useful commit messages. The author contends the structure optimizes for changelog automation at the expense of human understanding.

It’s the sort of process-religion debate HN loves, and the thread is predictably divided between people who credit the convention for tidy histories and automated releases and those who find it bureaucratic theater. Underneath is a real question about what commit history is even for.