📦 getgaal/gaal

gaal: One YAML to Sync Every AI Coding Agent

Reconcile every AI coding agent on your machine with a single command. If you juggle tools like Claude Code, Cursor, or Codex, you know the pain of manually copying skills and server settings everywhere they drift. Gaal solves this by acting as a central controller that reads one simple YAML file and pushes those updates to any agent you have installed. The standout feature is its auto-detection system. It scans your computer, finds seventeen different coding agents, and ensures every one of them has the exact same skills, files, and configurations. You can even run a dry-run to preview changes before anything is written, keeping your workflow safe and predictable.

🆔 @hackernewsgithubprojects

📦 dragthelake/ambient-context

Ambient Context: Your Mac Writes Its Own Diary

Your Mac quietly writes a diary of your workday so an A.I. assistant can actually remember what you did. Ambient Context is a tiny menu bar app that reads the text from your open windows and saves it to simple files on your computer. No screenshots, no cloud, just plain text you own. It’s designed to help A.I. tools answer questions like “what was I working on Tuesday?” by giving them a clean, private record of your day. If you’ve ever felt like your digital brain is leaking into the void, this is the fix.

📰 https://news.ycombinator.com/item?id=49429095

🆔 @hackernewsgithubprojects

📦 zju-real/agent-g2

Agent-G2: Smarter AI Learning Without Extra Costs

Agent-G2 reaches ninety-five percent success on complex AI tasks by figuring out exactly how much help each task needs during training. It solves a major problem in reinforcement learning where AI agents often get lost in long, multi-step tasks because rewards are rare. Instead of using a fixed amount of guidance or wasting compute on extra trial-and-error runs, the system uses a statistical method to find the sweet spot for help per task. This approach allows the AI to learn effectively without the high cost of additional testing. It is a clever, efficient way to teach machines to handle long, complex goals, making the learning process significantly more efficient.

🆔 @hackernewsgithubprojects

📦 jeffhajewski/latticedb

LatticeDB: The All-In-One Local Database

LatticeDB is the local database that finally puts graph connections, semantic search, and text search into a single portable file. It solves the complexity of building AI apps by letting you run a Cypher query that finds similar documents, traverses relationships to their authors, and filters by keyword, all in one go. The coolest part is that it is built in Zig for raw speed, delivering sub-millisecond search on a million vectors while keeping everything on your machine with zero server setup. You get the power of a cloud database in one tiny file, making it perfect for building local knowledge tools that feel instant and private.

📰 https://news.ycombinator.com/item?id=49437049

🆔 @hackernewsgithubprojects

📦 dualverse-ai/station

Station: The AI Lab That Finds New Math

Station is the open-world simulation that lets groups of artificial intelligence agents conduct real scientific research without human intervention. The project models a miniature research lab where different AI models act as independent scientists, spending time on tasks like proving complex geometric theorems or optimizing physical structures. It works best when a problem can be scored clearly and solved in a few hours, allowing the agents to iterate quickly and share their findings through a digital archive. One of the most surprising results was the system independently rediscovering a famous counterexample in mathematics within a single day.

🆔 @hackernewsgithubprojects

📦 thu-ml/turbodiffusion

TurboDiffusion: AI Video in 2 Seconds

A single consumer-grade graphics card can now render a five-second AI video in just two seconds, cutting the original processing time of over three minutes by nearly one hundred times. This tool, called TurboDiffusion, works by drastically reducing the number of steps the AI needs to build each frame, while also speeding up how it calculates visual details. It turns a tedious wait into something almost instantaneous, making real-time video generation a reality for individual creators. If you have ever waited around for an AI video to render, this is the shift that makes that wait feel like a ghost from a previous era.

🆔 @hackernewsgithubprojects

📦 mingxuzhang2/emrb

EMRB: The Benchmark for Signal Reasoning

The biggest AI models are failing at a basic engineering task, but not because they lack knowledge. EMRB, a new benchmark from researchers at KDD, tests if large language models can reason directly over raw electromagnetic signals. Instead of giving a model a neat spreadsheet of data, it hands over messy, raw audio-like capture files and asks for specific measurements. The model has to write code, run it in a sandbox, and figure out what is actually in the signal. Across fourteen major models, performance dropped from near eighty percent on simple tasks down to just twenty-one percent on complex system design.

🆔 @hackernewsgithubprojects

📦 kyuz0/pi-bench

How Pi-Bench Exposes AI Coding Agent Weaknesses

Pi-Bench gives you a clear, objective score for how well an AI coding agent actually handles real software bugs. Instead of guessing, it drops the agent into a sandboxed environment, lets it attempt a fix, and then runs the official test suite to see if the code works. The most revealing part is when it catches the agent stuck in a loop. By comparing high-precision cloud models against locally quantized versions, the tool showed that compressed models often repeat the same commands over and over. They forget what they just read and get stuck in a cycle of useless actions.

🆔 @hackernewsgithubprojects

📦 worldbench/diffusionopsd

DiffusionOPSD: Smarter Image Training

DiffusionOPSD is the training method that makes AI image generators significantly faster and more consistent. Standard models often struggle to learn from feedback because the final score only appears after the image is completely finished, leaving a gap in the learning process. This project fixes that by creating clear, intermediate targets for the model to follow during generation. It works like a smart tutor that gives instant, specific corrections at every step, not just at the end. The result is a system that reaches peak quality much faster than previous methods, using up to sixty percent less computing power.

🆔 @hackernewsgithubprojects

📦 simonsays1980/rl-triton

RL-Triton Speeds Up AI Credit Assignment

Boost your reinforcement learning pipelines by replacing slow, standard calculations with lightning fast GPU code. R L Triton is a specialized library that writes highly optimized kernels to handle the messy math of teaching machines how to act. It takes complex tasks like figuring out which past actions led to a reward and crunches them directly on your graphics processor. This speeds up training by two to three times compared to typical Python tools. By handling the heavy lifting of credit assignment more efficiently, you get faster results without rewriting your entire system. It is a smart way to make your AI models learn quicker and more accurately.

🆔 @hackernewsgithubprojects

📦 riii111/sabiql

Sabiql: The Database Tool That Doesn't Need Drivers

Sabiql is a terminal database client that completely bypasses the usual setup headache. Instead of wrestling with configuration files or installing heavy database drivers, it simply speaks directly to your existing Postgres or SQLite command line tools. This driver-less approach means zero dependency conflicts and a tool that works exactly as fast as your underlying system allows. The interface is built for speed, featuring a Vim-style workflow that keeps your hands on the home row while you browse schemas or run complex queries.

🆔 @hackernewsgithubprojects

📦 spark-to-paper-skills/spark-to-paper-skills

Spark-to-Paper-Skills: From Idea to PDF in One Prompt

Spark-to-Paper-Skills turns a single line of text into a fully compiled research paper for about ten dollars. It doesn’t just write text; it plans experiments, runs them, generates editable vector figures, and checks every citation against real records. The biggest surprise is the figure engine, which takes a rough AI-generated image and redraws it as clean, editable vector graphics that look professional and actually work. This tool solves the problem of turning a vague idea into a publishable draft without the endless back-and-forth of manual editing. It’s perfect for researchers who want to test their ideas quickly and see a real paper structure emerge.

🆔 @hackernewsgithubprojects

📦 thinkoffapp/carwatch

Your Car Joins Your Chat as a Real Agent

CarWatch is the garage project that turns your car into a full chat-room teammate. It plugs a small computer into your vehicle and a dashcam to create a digital agent that speaks for your car. Instead of just recording video, this agent joins your private group chat and posts updates like a colleague. It tracks when you leave home, reports battery levels, and alerts you if the car spots a problem while parked. The coolest part is that it runs entirely offline on a Raspberry Pi, so it never needs the cloud. It even answers questions you ask it by voice, all while keeping your data private.

📰 https://news.ycombinator.com/item?id=49435675

🆔 @hackernewsgithubprojects

📦 zhikaixu24/das

DAS: AI Writes Real Academic Surveys

Generate a full, publication-ready academic survey in under an hour. That is exactly what the Deep Academic Survey project does. It takes a research topic and automatically writes a complete paper, complete with organized sections, proper citations, and formatted figures. What makes it truly interesting is that it does not just summarize text; it builds a manuscript that looks like a human wrote it. The system checks its own work, fixing mistakes until the final document is solid. It even creates a benchmark to prove how well it performs against other tools. This is a big step toward automating the heavy lifting of academic research.

🆔 @hackernewsgithubprojects

📦 spronta/crawlie

Crawlie: The Free SEO Crawler for Humans and AI

Crawlie crawls your entire website to catch technical search engine optimization bugs and AI visibility issues that usually hide in plain sight. It runs over fifty checks directly in Rust, so it is blazing fast and works even on massive sites without crashing your computer's memory. The coolest part is that it drives a real headless browser, meaning it actually sees your site after JavaScript loads, just like a human does. This tool is built for both humans and coding agents, giving you plain English explanations for every fix. It is completely free, open source, and ready to make your content visible to both search engines and AI.

📰 https://news.ycombinator.com/item?id=48592731

🆔 @hackernewsgithubprojects

📦 openspace-lab/space-mining-with-robotics-list

Mining Mars with Robots: The Complete Research List

Explore how we mine beyond Earth with space robots, covering the full pipeline from remote sensing to physical extraction. This repository acts as a comprehensive guide to the field, organizing dozens of academic papers into a clear, step-by-step roadmap. It breaks down the technical journey into distinct levels: first, identifying targets using satellite data, and then deploying autonomous machines to navigate and dig. You will find specific sections dedicated to the mechanical designs of these rovers, their sensors, and the complex control systems that keep them stable on uneven ground. It also details the navigation tech, like visual mapping, that allows these machines to know exactly where they are.

🆔 @hackernewsgithubprojects

📦 evi-group-scu/v-zero

Teaching AI to See Details Without the Cheat Sheet

V-Zero teaches AI models to spot tiny details in images without needing a human to mark the correct answer. Usually, training a model to focus on specific parts of a photo, like a license plate or a small object, requires expensive, manual labeling. This project solves that problem by having a teacher model compare the same reasoning path against two different crops of the image: one that matters and one that doesn't. By seeing which crop actually supports the model's logic, it learns exactly where to look.

🆔 @hackernewsgithubprojects

📦 castorini/cmass

Testing AI on a 400 Billion Token Library

A new benchmark called ClimbMix Agentic Search Suite forces AI agents to hunt for answers inside a library of four hundred billion tokens. This massive dataset contains five hundred fifty three million documents, creating a realistic environment where finding information is genuinely difficult. The project takes existing questions and breaks them into small pieces to ensure the answers actually exist in this giant archive. Researchers found that even top-tier models struggle here, needing far more search calls to succeed. This tool helps developers test if their AI can truly dig through real world data instead of just memorizing facts.

🆔 @hackernewsgithubprojects

📦 wangqinsi1/spyrl

SpyRL: The Game That Trains AI to Write Better

SpyRL turns AI training into a high-stakes game of Who Is the Spy. Instead of relying on strict math problems, this project pits language models against each other in social deduction tasks. One player is secretly the spy, missing pieces of information, while others try to spot the fakes. This clever setup creates automatic rewards without needing human labels, allowing models to improve at creative writing and summarization. By gamifying the process, it solves the problem of teaching AI on open-ended tasks where answers are subjective. The result is models that get significantly better at storytelling and reasoning, proving that competition can drive real progress.

🆔 @hackernewsgithubprojects

📦 lykelee/cldefocus

Cldefocus Fixes AI Blurring

A new tool called Cldefocus solves a hidden weakness in AI deblurring by creating fake blurry photos that look exactly like real ones. Instead of relying on limited camera captures, this pipeline uses physics-based math to simulate how light actually bends through hundreds of different lenses. The result is a massive collection of realistic blurry images paired with their sharp originals. Models trained on this data learn to fix blur better than ever, even when they face cameras they have never seen before. The takeaway is that synthetic data, built on real optical physics, helps artificial intelligence handle the messy, out-of-focus photos of the real world.

🆔 @hackernewsgithubprojects