Top Stories

Mistral Large 4

1837 points · mistral.ai

France’s Mistral is back with its new flagship, Mistral Large 4, and the HN community rocketed it to the top of the front page with more than 1,800 points. Every new Mistral release doubles as a referendum on whether an independent European lab can stay in the same conversation as OpenAI, Anthropic, and Google — and this level of enthusiasm suggests it landed.

The comments run the usual gauntlet: benchmark scrutiny, questions about whether the weights are open or API-only, and debate over licensing and pricing. Mistral has long courted developers who want models they can actually self-host, so where this one falls on the open-vs-closed spectrum is the thread people keep returning to.

Sharing AI progress in mathematics

952 points · openai.com

OpenAI put out an update on applying its models to research-level mathematics, and HN turned out in force. Math is an unusually clean proving ground for the “can these things actually reason” question — a proof is either correct or it isn’t, with much less room for the hand-waving that clouds other benchmarks.

The discussion splits predictably between people impressed by verifiable results and skeptics asking how much is genuine novel contribution versus sophisticated retrieval and assistance. Either way, it’s the kind of story that sharpens the ongoing debate about what counts as reasoning.

EmbeddingGemma 2

340 points · blog.google

Google released EmbeddingGemma 2, an open, lightweight, multimodal embedding model. Embeddings are the unglamorous backbone of modern search and retrieval-augmented generation, so a small open model that handles multiple modalities is immediately practical for anyone building local or on-device RAG pipelines.

HN developers care because “small and open” means they can run it without a cloud dependency or per-token bill — and the multimodal angle hints at search over images and text with one model.

Decisions API is in public beta

312 points · developers.openai.com

OpenAI opened its Decisions API to public beta, pitching a structured primitive for getting models to make and justify discrete choices rather than just emit free-form text. For anyone building agents, reliable decision-making is the hard part — so a first-party API aimed squarely at it drew a lot of attention.

It’s worth reading alongside the Strands Decider model trending the same day; between them, “decisions as a dedicated building block” is clearly a theme the ecosystem is converging on.

OpenTPU – an open-source AI accelerator, developed by AI

292 points · github.com

This project is a double hook for HN: an open-source AI accelerator design, and one whose author says was itself developed with heavy AI assistance. Open hardware already draws a crowd here, and the meta-story of AI helping design the chips that run AI is exactly the kind of recursive twist the community loves to pick apart.

Expect comments weighing how far the design actually goes versus the headline, plus the usual deep-dive from people who do RTL and silicon for a living.

AnyPS5: port PS5 binaries to PC without emulation

281 points · github.com

AnyPS5 claims to run PlayStation 5 binaries on PC without emulation by mapping the console’s system libraries directly — reportedly around 87% of them so far. Reverse-engineering feats like this are catnip on HN, where the technical ingenuity of making one platform’s binaries run natively on another is appreciated on its own terms.

The thread carries the inevitable undercurrent about legality, DRM, and how Sony might respond, but the engineering is the draw.

Tell HN: GitHub refuses to remove cracked copies of my software after a month

273 points · news.ycombinator.com

A developer vented that GitHub still hasn’t taken down cracked copies of their paid software a month after reporting them. It struck a nerve because it sits at the intersection of two things HN cares deeply about: the livelihood of small indie developers and the responsibilities (and inconsistencies) of the platforms they depend on.

The comments fill with others’ takedown war stories, advice on escalation, and a broader argument about whether big platforms move fast enough on anything that isn’t a legal emergency for themselves.

Claude Code’s suggested message feature: the real customer is the model

211 points · zohaib.cc

A sharp little essay arguing that the “suggested message” feature in Claude Code is less about convenience for the user and more about shaping the data and feedback loop that improves the model. It’s the kind of product-design observation that makes HN readers sit up — reframing a UI affordance as a quiet data-collection mechanism.

Whether you buy the thesis or not, it sparked a good debate about incentives in AI tooling and who’s really being served by features that look like they’re for you.

Penguin Mail – open-source Rust email client for Linux with AI

192 points · penguin-mail.com

Penguin Mail checks nearly every HN box at once: it’s open source, written in Rust, built for Linux, and sprinkled with AI features. The Linux desktop email client is a perennially underserved niche, so a fresh, modern take always draws interest from people tired of the existing options.

Commenters dig into the AI integration (how much is genuinely useful versus bolted on), Rust’s suitability for a desktop app, and the practical questions of IMAP support and real-world performance.