Story cards stay below the sticky dock while audio, chapters, date, and brief navigation remain accessible.
Google dropped Gemini 3.5 Flash at I/O today, billing it as "frontier intelligence with action." The latest in Google's model family promises a leap in agentic capabilities, and with 338 points and nearly 300 comments on HN, the community is digging in hard. Details on benchmarks and pricing are emerging as developers test it against the current crop of frontier models.
French AI powerhouse Mistral is acquiring Emmi AI, an Austrian startup that builds physics-based AI models for industrial engineering across energy, automotive, semiconductors, and aerospace. Emmi raised a EUR 15 million seed round in 2025 — the largest ever for an Austrian startup — and brings 30+ researchers specializing in real-time simulations and digital twins. CEO Arthur Mensch says the deal "cements Mistral's leadership in industrial AI."
OpenAI is adopting Google DeepMind's SynthID watermarking system for AI-generated images and releasing a verification tool for content provenance. This cross-industry collaboration marks a significant step toward solving the deepfake attribution problem — if the two biggest AI image generators agree on a watermarking standard, it could become a de facto industry norm.
Antoine Zambelli (AI Director at Texas Instruments) open-sourced Forge, a Python framework that wraps small self-hosted LLMs with guardrails, context management, and synthetic tool injection to dramatically boost agentic task reliability. The top config — Ministral-3 8B on llama-server — hits 86.5% accuracy across 26 scenarios, narrowing the gap with frontier APIs. The research covers 97 model configurations and will be presented at ACM CAIS '26.
Also out of I/O 2026, Google is overhauling its core search experience around AI agents, describing it as "a new era for AI Search." The update brings together traditional search engine capabilities with agentic AI features, signaling that the search box is no longer just a query input — it's becoming an AI-powered task execution interface.
Apple announced a sweeping set of accessibility features powered by Apple Intelligence, coming later in 2026. Highlights include natural-language voice control (say what you mean instead of memorizing commands), AI-generated image descriptions for VoiceOver, automatic subtitles for uncaptioned video, and Vision Pro wheelchair control via eye tracking. All processing happens on-device for privacy.
Minnesota has become the first U.S. state to outright ban prediction markets, a move that lands squarely in the crosshairs of fintech startups like Polymarket and Kalshi that have been riding a wave of mainstream adoption. The ban signals growing regulatory pushback against event contracts, and could embolden other states to follow suit.
Andrew Nesbitt catalogs the many ways open source projects silently die — from ghost maintainers and burnout plateaus to hostile takeovers (the xz backdoor gets a nod) and lost publish credentials. The core warning: lockfiles perpetuate dead dependencies invisibly, and downstream consumers bear hidden risk from software they assume is still maintained.
A passion project 20 years in the making: Andrew Warkentin built a single Linux VM containing over 1,700 pre-installed operating systems spanning from the 1948 Manchester Baby to modern distributions. It covers 570 distinct OSes across 250+ platforms — mainframes, Unix variants, home computers, mobile — all one-click launchable with snapshot support. A stunning act of software preservation.
Tesla's $1 billion lithium refinery in Robstown, Texas — North America's first commercial-scale spodumene-to-lithium-hydroxide facility — is dumping 231,000 gallons of treated wastewater daily into a drainage ditch flowing toward Baffin Bay. Independent lab tests found hexavalent chromium, arsenic, and elevated heavy metals not covered by Tesla's discharge permit. The local drainage district has demanded Tesla halt discharge.
