Top Stories

Microsoft and OpenAI end their exclusive and revenue-sharing deal

663 points · bloomberg.com

The headline news of the day, and a structural shift in the AI industry: Bloomberg is reporting that Microsoft and OpenAI have unwound the exclusive cloud and revenue-sharing arrangement that has defined their relationship since 2019. The 600-comment thread is the place to read for what this actually means in practice — OpenAI is now free to negotiate compute deals with Oracle, Google, and AWS at scale, and Microsoft is no longer obligated to share Azure-derived revenue back to OpenAI. Commenters are split on who wins: the consensus take is that OpenAI gets the optionality it has been visibly straining against for two years, while Microsoft gets to commercialize its own frontier models (Phi, MAI) without the awkwardness of competing with its largest investee. Either way, the de-facto Microsoft-as-OpenAI’s-only-cloud era is over.

GitHub Copilot is moving to usage-based billing

463 points · github.blog

GitHub announced the long-rumored shift away from flat-rate Copilot pricing to metered usage, with the 363-comment thread doing what HN does best — picking apart the pricing math. The headline change: agentic features (Copilot Workspace, the new Copilot CLI agent) will be billed on completed tasks, while inline completions stay on a generous monthly allowance before metering kicks in. Reactions split predictably — heavy users alarmed at unpredictable bills, light users pleased at lower entry pricing, and a sharp middle current pointing out that this brings GitHub’s pricing in line with how Anthropic and Cursor already charge, which is itself a sign that the “AI dev tools as flat SaaS” era is ending. Worth reading if you administer Copilot for a team.

4TB of voice samples stolen from 40k AI contractors at Mercor

398 points · oravys.com

A serious breach disclosure: a security researcher reports that roughly 4TB of voice training data — including raw samples from about 40,000 contractors who do RLHF and voice work for Mercor’s labelling pipeline — has been exfiltrated and is circulating on grey-market data brokers. The 151-comment thread is where the practical risk gets unpacked: voice samples at this scale and quality are exactly the input you need for high-fidelity voice cloning and KYC bypass against bank phone trees, and the contractors who provided them did so under contracts that almost certainly didn’t anticipate this threat model. A live story to watch — Mercor hasn’t formally responded as of this writing, and the legal exposure for AI labelling vendors is suddenly much sharper.

Pgbackrest is no longer being maintained

381 points · github.com

A quiet but consequential announcement on the project’s GitHub: pgBackRest, the de-facto standard backup tool for production PostgreSQL, has been formally archived after its lead maintainer announced he no longer has time or sponsorship to keep it going. The 196-comment thread is essential reading if you run Postgres at any scale — half is operators sharing what they’re moving to (pg_basebackup + WAL-G is the leading suggestion, with Barman in second), and half is a sober discussion of how a tool this critical to the Postgres ecosystem ended up dependent on a single under-resourced maintainer. A useful prompt to audit your own backup posture this week.

China blocks Meta’s acquisition of AI startup Manus

227 points · cnbc.com

Beijing’s State Administration for Market Regulation formally blocked Meta’s proposed acquisition of Manus, the Chinese-origin AI agent startup that went viral last year, citing antitrust and “data sovereignty” concerns. The 134-comment thread is mostly geopolitical analysis: this is the first major Chinese block of a US AI acquisition, and the reasoning will set the template for how future cross-border AI deals get evaluated. A useful subthread digs into Manus’ actual technical position — they were one of the few non-US-headquartered teams shipping a competitive general-purpose agent, and Meta clearly wanted the team more than the IP. The deal’s collapse is a real datapoint for anyone modeling the bifurcation of the global AI stack.

Show HN: OSS Agent topped TerminalBench on Gemini-3-flash-preview

278 points · github.com

A genuinely impressive Show HN: an open-source coding agent (Dirac) that the author claims tops the TerminalBench leaderboard when paired with Google’s new Gemini-3-flash-preview — beating closed-source agents from the major labs at a fraction of the inference cost. The 110-comment thread is the substantive part: people running Dirac on their own benchmarks, debating whether TerminalBench is still a meaningful eval given how much agent harnesses now optimize for it specifically, and a useful side-discussion on the architectural choices (deterministic tool routing, no LLM-as-planner) that the author credits for the result. Worth cloning if you’re building agentic developer tooling.

”Why not just use Lean?”

236 points · lawrencecpaulson.github.io

Lawrence Paulson — the architect of the Isabelle proof assistant — writes a careful, generous essay answering the question he gets asked at every conference: why hasn’t the formal-methods world converged on Lean now that Lean 4 and mathlib have so visibly won the mindshare? The 155-comment thread is the kind of high-quality language-design argument HN occasionally produces, with serious researchers from the Lean, Coq/Rocq, and Isabelle camps all weighing in on tradeoffs around tactic languages, automation, and what mature library ecosystems actually require. Required reading if you’ve been following the formal-methods renaissance.

Quarkdown — Markdown with Superpowers

226 points · quarkdown.com

A new Markdown extension with serious ambitions: Quarkdown adds a Pandoc-class output pipeline (HTML, LaTeX, slides, books) on top of a typed templating layer that lets you embed real expressions and reusable components without leaving the document. The 71-comment thread is the usual Markdown-extension skepticism — “we already have MDX/Typst/Asciidoc” — but with a useful sub-current of writers and academics describing what specifically they could do with Quarkdown that they couldn’t with the incumbents. Whether it sticks or not, the design choices are worth a look if you write technical content for a living.

Networking changes coming in macOS 27

165 points · eclecticlight.co

Howard Oakley’s careful walkthrough of the networking-stack changes Apple is shipping in macOS 27 (currently in developer beta) — most notably, the long-deprecated pf firewall is finally being replaced by a new userland-configurable filter framework, and several long-standing oddities around private-relay and VPN interaction get tidied up. The 151-comment thread is the practical part: developers and sysadmins flagging what’s going to break, what won’t, and which third-party VPN/firewall vendors have already shipped compatibility builds. If you ship Mac software that touches the network, this is the post to bookmark before WWDC.

The woes of sanitizing SVGs

151 points · muffin.ink

A sharp engineering write-up from a developer at Scratch (the kids’ coding platform) on what it actually takes to safely accept user-uploaded SVG art at scale. The 58-comment thread is full of war stories from people who have tried this and learned the hard way that SVG is essentially “HTML with a different file extension” — full of script execution, foreignObject embedding, and CSS-driven exfiltration vectors that no off-the-shelf sanitizer cleanly handles. A genuinely useful reference if your product accepts user SVG, and a quiet argument for rasterizing first and asking questions later.