Published on May 20, 2026 at 06:00 CEST (UTC+2)
Railway Blocked by Google Cloud (275 points by aarondf)
This status page from Railway reports a significant incident where Google Cloud blocked their services, impacting deployments and network operations across multiple regions. The page shows historically high uptime (99.98% overall) with brief dips in March and April, but current May metrics indicate 100% uptime for all services including US East, Southeast Asia, and EU West. The incident appears to have been resolved, as the dashboard now shows full operational status across all Railway Metal and Edge Network locations.
Gemini 3.5 Flash (642 points by spectraldrift)
Google announces Gemini 3.5 Flash, a new family of models that combine frontier intelligence with agentic capabilities, available globally to billions through the Gemini app, Google Search AI Mode, and developer platforms like Google Antigravity and AI Studio. The model excels at complex, long-horizon tasks for agents and coding, and is the first in a series that will also include a more powerful 3.5 Pro version currently used internally. Key executives including Koray Kavukcuoglu, Jeff Dean, and Noam Shazeer highlight this as a major leap toward building more capable, intelligent AI agents.
Ben Welsh made an index of all FiveThirtyEight articles on the Internet Archive (57 points by ChocMontePy)
Ben Welsh created fivethirtyeightindex.com, an archive that indexes 21,350 pages from FiveThirtyEight preserved by the Internet Archive, spanning from 2008 to 2025. The site allows browsing by year and byline, with Nate Silver contributing the most articles (4,966), followed by Neil Paine, Walt Hickey, and others. This makes the full historical record of FiveThirtyEight's data journalism easily searchable and accessible.
I’ve built a virtual museum with nearly every operating system you can think of (642 points by andreww591)
The Virtual OS Museum is a pre-configured Linux VM containing hundreds of operating systems and standalone applications running under emulation, from the Manchester Baby (1948) to modern systems. It includes a custom launcher with snapshot functionality to recover broken installations, and covers mainframes like CTSS and MVS, minicomputers like TOPS-10 and Multics, and early Unix versions. The project aims to let users explore OS history without configuring emulators or worrying about corrupting installations.
Google changes its search box (419 points by berkeleyjunk)
Google announces the biggest upgrade to its search box in over 25 years, integrating Gemini 3.5 Flash as the default AI model in AI Mode globally. AI Mode has surpassed one billion monthly users since its debut, with query volume doubling every quarter and reaching an all-time high last quarter. The new intelligent search box leverages advanced AI to handle complex, hyper-specific questions and marks a continued transformation of Search into an agentic experience.
Remove–AI–Watermarks – CLI and library for removing AI watermarks from images (156 points by janalsncm)
This open-source CLI and library removes visible and invisible AI watermarks from images generated by models like Google Gemini, ChatGPT/DALL-E, Stable Diffusion, Adobe Firefly, and Midjourney. It strips SynthID, C2PA Content Credentials, EXIF/XMP "Made with AI" labels, and visible sparkle overlays in a single command. The repository has gained 321 stars and 24 forks, indicating community interest in circumventing AI content provenance measures.
OpenAI Adopts Google's SynthID Watermark for AI Images with Verification Tool (224 points by smooke)
OpenAI announces it is adopting Google’s SynthID watermarking technology for AI-generated images, along with a verification tool to help users identify AI content. This marks a rare collaboration between the two AI leaders on content provenance, aiming to create a standardized approach for marking and detecting synthetic media. The move signals growing industry consensus on the need for robust watermarking to combat misinformation.
Show HN: Forge – Guardrails take an 8B model from 53% to 99% on agentic tasks (324 points by zambelli)
Forge is a Python framework that dramatically improves the reliability of self-hosted LLM tool-calling by using guardrails such as rescue parsing, retry nudges, and step enforcement, along with VRAM-aware context management. The framework lifts an 8B local model (Ministral-3 8B Instruct Q8) from 53% to 99% on multi-step agentic evaluation tasks across 26 scenarios. This demonstrates that small, local models can achieve near-top-tier performance when paired with robust reliability layers.
The Mercury logic programming system (42 points by Antibabelic)
The Mercury logic programming system is a long-established open-source project on GitHub with over 25,000 commits and 1,000 stars, featuring a compiler, runtime, and extensive documentation. It is a functional logic programming language with strong typing and efficient execution, used for applications requiring declarative reasoning. The repository remains actively maintained with build scripts, test suites, and Docker support.
Mistral AI acquires Emmi AI (190 points by doener)
Mistral AI acquires Emmi AI, a European Engineering AI company specializing in Physics AI models for industrial simulation, in one of Europe's most strategic AI acquisitions. Emmi’s team of over 30 researchers and engineers will join Mistral’s Science and Applied AI teams, and Linz, Austria will become an official Mistral AI office. The acquisition aims to create the leading AI stack for industrial engineering, targeting sectors like energy, automotive, semiconductors, and aerospace.
1. Agentic AI is becoming the dominant paradigm for model deployment Both Gemini 3.5 Flash and the Forge framework explicitly focus on agentic workflows—executing complex, multi-step tasks with tool calling and orchestration. This shift from simple Q&A to autonomous, goal-driven systems represents a major architectural evolution. For developers, investing in frameworks that provide guardrails, retry logic, and context management will be critical to making agents reliable in production.
2. Watermarking and provenance are escalating into an arms race OpenAI adopting Google’s SynthID while a separate open-source tool (Remove-AI-Watermarks) immediately emerges to strip those watermarks highlights a cat-and-mouse dynamic. As AI-generated content proliferates, content provenance will become a regulatory and trust battleground. Companies must plan for both watermarking deployment and the inevitable countermeasures, and should consider layered approaches (e.g., invisible + metadata + visible marks) to increase removal difficulty.
3. Small, local models with guardrails can rival massive cloud models on specific tasks Forge demonstrates that an 8B parameter local model, when enhanced with rescue parsing and step enforcement, can achieve 99% accuracy on agentic benchmarks—performance that typically requires much larger models. This trend lowers the barrier for on-device and privacy-sensitive AI deployments. Enterprises should explore "small model + reliability layer" stacks as a cost-effective alternative to API-dependent large models.
4. AI is transforming search into an agentic, conversational experience Google’s search box upgrade, powered by Gemini 3.5 Flash, redefines the classic text-input paradigm into an interactive AI agent that handles complex, multi-step queries. With AI Mode surpassing 1 billion users, this signals a permanent shift in how users interact with data. Search engines, knowledge bases, and product catalogs all need to become "agent-ready"—supporting tool-calling, context retention, and iterative reasoning.
5. Industrial and physics AI is a major acquisition frontier Mistral’s acquisition of Emmi AI for industrial engineering underscores the growing importance of domain-specific AI in manufacturing, energy, and aerospace. Physics-informed neural networks and simulation-accelerating models are becoming strategic assets. Expect more consolidation as generalist AI companies acquire vertical specialists to offer full-stack industrial transformation, particularly in Europe and the US.
6. Cloud infrastructure reliability remains a critical concern for AI services The Railway incident, where Google Cloud blocked services, shows that even highly available platforms (99.98% uptime) can face catastrophic outages due to cloud provider dependencies. As AI workloads become more mission-critical, multi-cloud or hybrid architectures will become essential for resilience. Teams should formalize failover strategies and avoid single-vendor lock-in for compute and storage.
7. Preservation and indexing of AI-generated data is gaining attention The FiveThirtyEight index and the Virtual OS Museum both highlight a growing cultural and technical need to preserve digital artifacts—whether they are human-written articles or historical software. As AI generates vast amounts of content, systematic archiving and searchable indexes will become vital for accountability, research, and historical record. This parallels the need for provenance tools and may drive new standards for data lineage.
Analysis generated by deepseek-reasoner