Top Stories

Claude: System Prompts

672 points · platform.claude.com

Anthropic published the actual system prompts it ships with Claude, and the HN crowd pounced. System prompts are usually a closely guarded piece of the secret sauce — the invisible instructions that shape a model’s tone, refusals, and behavior — so seeing them laid out in official release notes is a rare bit of transparency. Expect people to comb through every line for clues about how the model is steered.

The discussion is the usual mix of appreciation for the openness and nitpicking over specific instructions. For anyone building on top of these models, the prompts are a practical reference for what the base assistant already assumes so you don’t fight it in your own layer.

Qwen 3.8 27B Is Excellent, but It Defaults to Overthinking Things

528 points · simonwillison.net

Simon Willison’s hands-on take on Alibaba’s latest Qwen release struck a nerve. His verdict: the 27B model is genuinely strong for its size, but it has a habit of burning tokens on elaborate chains of reasoning even for simple prompts. That “overthinking” tendency is becoming a recurring theme as reasoning-tuned models proliferate — great for hard problems, wasteful for easy ones.

HN loves Willison’s model write-ups because they’re concrete and reproducible, and the comments dig into how to rein in the reasoning behavior and how open-weight Chinese models keep closing the gap with frontier labs.

Stripe Will Reportedly Acquire OpenRouter for $7B+

371 points · techcrunch.com

Stripe is reportedly buying OpenRouter, the popular AI gateway that lets developers route requests across dozens of LLM providers through one API, for north of $7 billion. It’s a striking price for a company that essentially sits as a billing-and-routing layer in front of the model providers — which is precisely why a payments giant like Stripe would want it.

The deal reframes the “picks and shovels” thesis of the AI boom: the money may flow less to any single model and more to whoever owns the metering, routing, and payment rails. Commenters are debating whether this cements Stripe as the toll booth of AI inference or overpays for a thin margin business.

A Third World Engineer Responds to “RISC-V: They Should Have Known Better”

527 points · rvembedded.com

A pointed rebuttal to a widely shared critique of the RISC-V instruction set, written from the perspective of an embedded engineer outside the usual Western tech centers. The author argues that many complaints about RISC-V’s design miss the point of why an open, royalty-free ISA matters for engineers who can’t simply license ARM.

It resonated because it reframes an architecture debate as an access-and-economics debate. The comments split between ISA purists arguing technical merits and folks championing RISC-V’s role in democratizing chip design.

Tell HN: Cloudflare Silently Injects Its Analytics When You Switch Nameservers

526 points · news.ycombinator.com

A user reports that pointing your domain’s nameservers at Cloudflare can result in Cloudflare’s Web Analytics script being injected into your pages without an explicit opt-in. For a privacy-sensitive crowd, silent script injection touches a nerve regardless of how benign the analytics are.

The thread is a lively debate about defaults, consent, and how much a CDN/DNS provider should be doing to your traffic behind the scenes. Even fans of Cloudflare tend to agree that “silent” and “injection” are not words you want near your infrastructure.

Anthropic’s “Watermark” Text Adulteration in Claude Is a Perversion of Writing

316 points · daringfireball.net

John Gruber takes aim at reports that Claude’s output includes invisible watermarking — subtle character-level tweaks meant to flag AI-generated text. His argument: quietly altering the characters of someone’s writing to embed a signal corrupts the text itself, and users deserve to know.

It’s a sharp counterpoint to the growing push for AI provenance and detectability. The comments wrestle with the real tension: watermarking could help fight AI slop and fraud, but doing it invisibly to text people consider their own crosses a line for many.

The AI Credit Resale Economy

297 points · vectoral.com

A look at the emerging gray market of “token brokers” — intermediaries who buy AI inference credits in bulk and resell access, arbitraging pricing tiers, regional availability, and rate limits. As model access becomes a commodity, a whole shadow economy is forming around reselling it.

HN readers find this fascinating because it mirrors older markets (cloud compute reselling, API arbitrage) and raises questions about terms-of-service enforcement and where the margins really live in the AI stack.

Protobuf Has LSP Support

157 points · buf.build

Buf announced a Language Server Protocol implementation for Protocol Buffers, bringing autocomplete, go-to-definition, and inline diagnostics to .proto files across any LSP-capable editor. Anyone who’s wrangled large protobuf schemas by hand will appreciate not flying blind.

Developer-tooling launches like this reliably do well on HN, and the comments cover editor integrations, how it compares to existing protobuf plugins, and wish-list features for schema-heavy codebases.

GIMP Development Update

192 points · gimp.org

The GIMP team’s latest progress report drew a warm reception from the open-source faithful. After the long road to the 3.x series, steady development updates are a good sign for the project’s momentum, covering new features, performance work, and the perennial goal of narrowing the gap with commercial editors.

The thread is the familiar GIMP conversation — genuine affection for the project mixed with UX critiques and comparisons to Photoshop and newer challengers.

Reticulum — Decentralized Mesh Network

137 points · reticulum.network

Reticulum is a cryptography-based networking stack designed to run over anything — LoRa radios, packet radio, serial links, or plain TCP/IP — without central servers or hierarchical addressing. It’s aimed at resilient, off-grid, censorship-resistant communication.

The project appeals to HN’s mesh-networking and self-reliance streak, and the comments explore real-world deployments, comparisons to other decentralized protocols, and how practical it is to build on today.