Published on July 10, 2026 at 06:00 CEST (UTC+2)
Show HN: Getting GLM 5.2 running on my slow computer (447 points by vforno)
Show HN: Getting GLM 5.2 running on my slow computer
This project, “colibri,” lets users run the massive GLM-5.2 model (744B parameters, Mixture-of-Experts) on a consumer machine with only 25 GB of RAM. It achieves this by streaming only the routed experts from disk while keeping the dense portion resident in memory, all in pure C with no dependencies. The tool demonstrates that even incredibly large MoE models can be made accessible on modest hardware through clever memory management.
EU Parliament greenlights Chat Control 1.0 (1056 points by rapnie)
EU Parliament greenlights Chat Control 1.0
The European Parliament approved a controversial measure allowing suspicionless mass scanning of private communications, despite a majority of voting MEPs opposing it. The regulation permits such scanning until 2028, with a symbolic exemption for encrypted messages. Civil rights advocates warn that this undermines democracy and fails to provide genuine child protection, marking a major setback for digital privacy.
Focus (75 points by iacguy)
Focus
An early Facebook engineer reflects on the brutal startup life, working 120-hour weeks with no hobbies and constant on-call duties. He emphasizes that the intense scarcity of resources and time forced the team to develop unwavering focus, saying no to non‑essential initiatives. The piece serves as a reminder that constraint can drive conviction and prioritization, even in a chaotic environment.
Ben Bernanke Joins Anthropic Oversight Trust (15 points by Jimmc414)
Ben Bernanke Joins Anthropic Oversight Trust
Former Federal Reserve Chair Ben Bernanke has been appointed to Anthropic’s Long-Term Benefit Trust, an independent body ensuring the company develops AI responsibly. Bernanke brings deep expertise in macroeconomics and crisis management, and his involvement highlights the growing recognition that AI governance requires experts from diverse fields. The move underscores Anthropic’s commitment to long-term societal benefit over short-term profit.
Show HN: 18 Words (869 points by pompomsheep)
Show HN: 18 Words
This is a daily word‑challenge game where players try to survive a puzzle using only 18 words. It includes features like hints, a paused mode, and an archive of past puzzles. The simple, engaging format appears designed for casual play and social sharing, with feedback options and a link to the creator’s other game.
GPT-5.6 (1126 points by logickkk1)
GPT-5.6
OpenAI announced a new iteration of its flagship model, GPT-5.6. While the full content is not available, the high score (1126 points) indicates strong community interest. The release likely brings improvements in reasoning, efficiency, or multimodal capabilities, continuing OpenAI’s rapid cadence of model updates.
Train sim created by just one person is being called the best ever made (353 points by oumua_don17)
Train sim created by just one person is being called the best ever made
Running Train, a Japanese train simulation developed by a solo developer, has been praised for its hyper-realistic graphics and meticulous attention to detail. Players can either take full control of speed and braking or simply watch the train run automatically. The game’s immersive atmosphere and single‑creator origin have captivated both train enthusiasts and general audiences.
Star Just Ate a Planet, and It's Not Done Yet (27 points by wglb)
Star Just Ate a Planet, and It’s Not Done Yet
Astronomers observed a star in the process of engulfing a planet, with the star still actively consuming the planetary material. The event offers rare direct evidence of planetary engulfment, a process that helps scientists understand the final stages of star and planet evolution. The ongoing nature of the event provides a unique opportunity for further study.
Hy3 (410 points by andai)
Hy3
Tencent released Hy3, a new AI model or framework (details sparse from the preview). Given the high score (410 points), Hy3 likely represents a significant advancement, possibly a large language model or a highly optimized inference engine. It signals Tencent’s continued investment in cutting‑edge AI research.
Postgres rewritten in Rust, now passing 100% of the Postgres regression tests (445 points by SweetSoftPillow)
Postgres rewritten in Rust, now passing 100% of the Postgres regression tests
The open‑source project pgrust has reimplemented PostgreSQL entirely in Rust, achieving full compatibility by passing over 46,000 regression queries. It is disk‑compatible with existing Postgres databases and was built with heavy use of AI‑assisted programming. The goal is to make Postgres easier to evolve and maintain while preserving its behavior.
Implications: Expect more open‑source projects that combine MoE architectures with disk‑backed inference, lowering the cost of deploying large models on edge devices or personal machines.
AI governance attracts top non‑tech experts
Implications: More AI companies will appoint prominent figures from finance, law, and environmental science to their oversight bodies, shaping public trust and regulatory engagement.
AI‑assisted software rewriting of critical infrastructure
Implications: Expect a wave of AI‑driven rewrites of other foundational software (e.g., SQLite, Redis, Linux utilities), potentially leading to more memory‑safe, efficient codebases.
Rapid model iteration: GPT-5.6 and Tencent Hy3
Implications: Organizations must plan for continuous model evaluation and integration. Smaller players may find it harder to keep up, while open‑source alternatives (e.g., Llama 4, Mistral) become more attractive.
Privacy vs. AI: regulatory friction is growing
Implications: Expect privacy‑preserving techniques (differential privacy, on‑device AI, federated learning) to become critical differentiators. Companies operating in Europe will need to navigate conflicting mandates between innovation and surveillance.
The power of extreme focus in AI product development
Analysis generated by deepseek-reasoner