Top Stories

1583 points · browsergate.eu

A bombshell privacy report reveals that LinkedIn runs hidden code to scan users’ installed browser extensions, identifying software that can expose religious beliefs, political views, disabilities, and job-seeking activity. Even worse, LinkedIn allegedly uses this data to map which companies use competing sales tools — essentially extracting competitor customer lists without anyone’s knowledge. The scanning happens silently via invisible tracking elements and none of it is disclosed in LinkedIn’s privacy policy. The HN community is furious, and this is drawing comparisons to some of the worst corporate surveillance practices. Expect regulatory fallout, especially in the EU.

Google Releases Gemma 4 Open Models

1205 points · deepmind.google

Google DeepMind dropped Gemma 4, a new family of open-weight models built on Gemini 3 research. The lineup spans four sizes — from tiny E2B and E4B models that run on phones and Raspberry Pis, up to 26B and 31B models with frontier-class performance on consumer GPUs. The 31B model posts impressive benchmark numbers (85.2% MMLU, 89.2% AIME 2026) while supporting native function calling, multimodal input (audio + vision), and 140 languages. This is a serious move in the open-weight model race and gives developers a lot of firepower without needing cloud APIs.

Sweden Goes Back to Basics: Swapping Screens for Books

762 points · undark.org

Sweden — once a poster child for digital-first education — is reversing course in dramatic fashion. The government allocated $83 million for physical textbooks and $54 million for fiction and non-fiction books, aiming to give every student physical texts for each subject. The pivot comes after years of declining standardized test scores and growing research showing that screen-based learning can hinder reading comprehension, especially for younger students. The HN discussion is a lively debate about screen time, deep reading, and whether this signals a broader global trend.

Qwen3.6-Plus: Alibaba’s Push Toward Real Agentic AI

447 points · qwen.ai

Alibaba released Qwen3.6-Plus, a model explicitly designed for agentic workflows — not just chatting, but planning, executing, and iterating on complex tasks autonomously. It features a 1-million-token context window, always-on chain-of-thought reasoning, native function calling, and up to 65K output tokens. The model is compatible with popular coding assistants like Claude Code and Cline. Alibaba is positioning this squarely at the “agentic coding” use case, and the benchmarks suggest it can actually deliver on that promise.

AMD’s Lemonade: Fast, Open Source Local LLM Server

459 points · lemonade-server.ai

AMD launched Lemonade, an open-source local AI server that auto-configures for your GPU and NPU hardware. The native C++ backend weighs just 2MB and supports chat, vision, image generation, transcription, and speech synthesis — all running locally. It’s OpenAI API-compatible out of the box, so it plugs straight into tools like Open WebUI and GitHub Copilot. For anyone wanting to run AI locally without cloud dependencies, this looks like the most frictionless option yet.

Decisions That Eroded Trust in Azure

382 points · isolveproblems.substack.com

A former Azure Core engineer published a detailed post-mortem on what went wrong inside Microsoft’s cloud division. The highlights are damning: 173 unaccountable management agents on each Azure node that nobody could explain, unrealistic hardware porting plans, and leadership that ignored warnings from experienced engineers. The piece alleges these failures contributed to damaged relationships with OpenAI and U.S. government customers. It’s a fascinating case study in organizational dysfunction at massive scale.

Cursor 3: A Completely New IDE for the Agent Era

308 points · cursor.com

Cursor ditched its VS Code fork and built an entirely new interface from scratch, centered around multi-agent workflows. The big features: a unified sidebar managing agents from mobile, web, desktop, Slack, and GitHub; seamless local-to-cloud handoff for agent sessions; and a redesigned diff review system. They’ve also launched a plugin marketplace. The vision is ambitious — Cursor wants to be the workspace where coding agents eventually drive themselves, with humans reviewing the output.

OpenAI Acquires TBPN Talk Show

164 points · openai.com

In a surprise move, OpenAI acquired TBPN (Technology Business Programming Network), the popular daily tech talk show hosted by John Coogan and Jordi Hays. The deal reportedly cost in the low hundreds of millions. TBPN will retain editorial independence and continue making their own programming choices. It’s OpenAI’s first media acquisition and signals a new strategy around direct audience engagement. The HN community is debating whether this is savvy brand-building or an unsettling convergence of AI and media.

Artemis II Computer Running Two Instances of Outlook — Nobody Knows Why

342 points · bsky.app

In the most delightful space-meets-software bug of the year, someone discovered that the Artemis II spacecraft’s computer system is running two instances of Microsoft Outlook — and NASA engineers reportedly can’t figure out why. The HN thread is a masterclass in engineering humor, with comments ranging from “have you tried turning it off and on again” to serious discussions about legacy software in mission-critical systems.