📦 radixark/miles

Miles: The RL Framework for Massive Models

Miles is the reinforcement learning framework that lets massive AI models learn without crashing. It connects the generation and training processes so they run at the same speed, which stops huge jobs from falling apart. The standout feature is fault tolerance. If a piece of hardware fails during a run, Miles fixes it and keeps going right where it left off. There is no pause and no restart. This makes training frontier models far more reliable. It is a powerful tool for anyone building serious AI. If you are training large models, this is the stability layer you need.

🆔 @hackernewsgithubprojects

📦 player-yn/pawwork_zhuazhua

Paw Work: The Chrome Extension That Turns Web Pages Into Real Files

Paw Work is the Chrome extension that turns the text and images you see on a web page into actual, editable office files. It solves the annoying problem of copy-pasting messy content from a browser into a blank document by letting you highlight exactly what you need and describe the result you want in plain language. You point at a product table, ask for a comparison sheet, and it builds a live spreadsheet right there, ready for you to tweak. The whole thing runs locally on your machine, so your data stays private and you control the AI keys.

🆔 @hackernewsgithubprojects

📦 superdesigndev/treg

Treg: The Universal Key for Agent Tools

Treg acts as a universal gateway for artificial intelligence agents, solving the messy problem of connecting them to dozens of different data sources. Instead of every agent needing its own unique keys and code for each service, Treg provides one shared connection point. This means your AI can access tools like web scraping, ad tracking, and business data through a single, standardized setup. It removes the friction of managing scattered credentials, making it easy for agents to find and use the information they need. Think of it as a master key that simplifies how smart programs interact with the outside world.

🆔 @hackernewsgithubprojects

📦 facebookresearch/jepa-wms

JEPA-WMs: The AI that learns physics

JEPA-WMs is the world model that learns physics directly from video instead of pixels. Meta AI built this tool to teach robots how the physical world works by watching footage of them moving. It does not just store images. It creates an internal map that predicts what happens next when you grab or push an object. This allows a robot to plan its moves before it actually does them. You get a smarter machine that understands cause and effect without needing constant human correction. It is a clear step toward robots that truly grasp how things move and interact in real life.

🆔 @hackernewsgithubprojects

📦 rlinf/rpent

RPent: The Robot Framework That Learns

A robot arm can now watch its own mistakes and fix them in real time. The project is called RPent. It is a framework for building agents that interact with the physical world. Instead of relying on one frozen model, it combines perception, reasoning, and memory. The agent plans an action, executes it, and reflects on the result. If it fails, it uses that memory to adapt and try again. This recursive loop lets the system improve beyond its initial design. It solves the problem of rigid robots that cannot handle new situations. It is interesting because it turns static software into a learning physical partner.

🆔 @hackernewsgithubprojects

📦 ath-maas/marco-deepresearch

Marco DeepResearch: The AI Agent That Actually Works

Marco DeepResearch exposes the frustrating gap between human experts and AI agents in real-world tasks. The project’s standout discovery is that while human tariff experts classify products with ninety-five percent accuracy, the best AI systems manage less than half that score. This massive difference highlights a critical limitation in how current models handle complex, hierarchical rules found in trade, law, and medicine. By building specific benchmarks and a self-evolving memory system, the team shows exactly where agents fail and how they can improve. The goal is not just to make chatbots smarter, but to create reliable workers that can handle messy, real-world information without constant human correction.

🆔 @hackernewsgithubprojects

📦 swarm-subnet/langostino

Langostino: The Drone You Can Actually Build

Langostino is the open-source drone that finally makes autonomous flight feel like a weekend hobby project rather than a black-box mystery. It’s a complete reference platform built with R O S and A I that hands you the exact parts list, wiring diagrams, and code needed to assemble a real, flying machine from scratch. The coolest part is that it isn’t just a simulation; you physically build the airframe, flash the flight controller, and send it outside to hold its position in mid-air. It cuts through the usual engineering fog by giving you a clear, step-by-step path from buying your components to seeing your own drone fly.

🆔 @hackernewsgithubprojects

📦 lmwilki/civ6-mcp

Civ6 MCP: Letting AI Play Civilization

Civ six M C P is the artificial intelligence project that finally lets language models play real games of Civilization VI. It connects directly to a running game, allowing the agent to move units, manage cities, and handle diplomacy through the game's own rules. No cheats and no screen reading are needed, just pure strategic reasoning over hundreds of turns. This makes it a perfect testbed for evaluating how well AI handles long-term planning and complex resource management in a dynamic environment. It is a fascinating look at how we can push language models beyond simple chat into deep, interactive gameplay.

🆔 @hackernewsgithubprojects

📦 tigerless-labs/influencer-discovery

influencer-discovery: The Zero-Dependency Outreach Engine

influencer-discovery is the zero-dependency pipeline that actually finds creators who bring their own audience. It runs as a skill for Claude Code, searching fifteen platforms to hunt down real people with established followings. The killer feature? It filters out people building their own products because they need reach, not your promotion. It only appends genuine contacts with public email addresses to your Google Sheet. There is no heavy database to install and no third-party bloat. It is pure Python standard library. You get a clean list of reachable influencers without the noise of self-promoters. It is the smart, quiet way to build your outreach list. Stop guessing who to contact.

🆔 @hackernewsgithubprojects

📦 0xohioism/polymarket-crypto-5m-15min-trading-bot

Polymarket Crypto Trading Bot

Polymarket Crypto Trading Bot is the public due-diligence overview that proves a live automated system actually works on short-term crypto markets. Instead of hiding behind backtests, this project shares real trading data from a live account that has traded over thirty thousand markets. It automates buying and selling on Polymarket for Bitcoin, Ethereum, and Solana price moves in five-minute windows. The most interesting part is the transparency. You can see the actual wins, the losing days, and the fees, so you know exactly what to expect. This isn’t a magic money printer, but it is a real, working system with verifiable results.

🆔 @hackernewsgithubprojects

📦 swufe-db-group/texlite

TexLite: The Self-Hosted LaTeX Editor

Write shared academic papers in your browser without handing your files to a cloud server. TexLite is a lightweight, self-hosted collaborative editor built for small research teams who want the live editing and commenting features of Overleaf, but with complete control over their data. The coolest part is how light it is; instead of running a massive, heavy container stack, it simply uses the LaTeX engine you already have installed on your computer. That means faster performance, zero external dependencies, and your documents stay safe on your own hard drive. It is the perfect fit for trusted groups who value privacy and speed.

📰 https://news.ycombinator.com/item?id=49450500

🆔 @hackernewsgithubprojects

📦 alchaincyf/huashu-excel

The Excel Audit Skill

Your AI agent is confidently lying to you about your sales numbers. The repository huashu-excel fixes this by making every single digit traceable to its exact source cell. It stops the common practice of blindly loading data into pandas, which silently mangles merged cells and duplicate rows. Instead, it performs a deep health check on the raw file to find dirty data. It then uses the totals rows already inside your spreadsheet as free checksums to verify every calculation. If your numbers do not match the original sheet, it flags the error immediately. This prevents the silent disasters where a wrong spreadsheet gets delivered as fact.

🆔 @hackernewsgithubprojects

📦 qufei1993/dsh-desktop

DSH Desktop: Zero Setup DeepSeek Harness

DSH Desktop removes the biggest hurdle to running the official DeepSeek Harness by bundling its own version of Node.js and a package manager directly inside the installer. This means you never have to hunt down the right runtime or configure your computer to get it working. The app launches the official interface in a secure, isolated window and includes a built-in manager to switch between different versions of the tool without touching your existing data. It is a clean, community-built wrapper for macOS and Windows that keeps the underlying software exactly as the developers intended, just in a package that is ready to run immediately.

🆔 @hackernewsgithubprojects

📦 flesymeb/harnessofharness

Harness of Harness: The AI That Builds Games Over Time

Harness of Harness is the autonomous framework that actually builds playable software over time. Instead of just writing code, it runs a multi-day process where an AI team plans, codes, and tests a project repeatedly. Each round improves the last one, letting the system fix mistakes and add polish gradually. It is cool because it produced real games, like a first-person shooter with a story and audio, entirely on its own. You watch the project grow from a rough idea into something you can actually play. It proves that AI can handle long, complex work, not just quick tasks.

🆔 @hackernewsgithubprojects

📦 2akouwu/reverify

Reverify: The AI That Doesn't Lie

When an A.I. analyzes a computer program, it usually just guesses what the code does, often inventing details that don't exist. Reverify fixes this by pairing a language model with a set of hard-working tools that act as a judge. The model proposes a theory about the binary, and the tools check that theory against the actual bytes. If the guess is wrong, the tools prove it is false. If it is right, the guess is verified. This creates a loop where the A.I. only reports facts that can be proven from the code itself, making the results trustworthy for real-world security work.

🆔 @hackernewsgithubprojects

📦 gbin/defragger

Defragger: The Linux Defragmenter You Actually Missed

Defragger lets you watch your Linux disk physically reorganize itself in real time, satisfying a deep itch for the old Windows block map. Built with Rust and Qt, it is not just a nostalgic wrapper, but a genuine tool that talks directly to the kernel to move file chunks. The coolest part? It runs the heavy lifting through a secure helper, so your main app never actually needs root access. It is a rare mix of visual satisfaction and serious utility for anyone curious about how files really live on your drive. If you want to see your storage breathe, this is the one to try.

📰 https://news.ycombinator.com/item?id=49438865

🆔 @hackernewsgithubprojects

📦 sharkspicy-nlp/ruleweaver

RuleWeaver: The New Standard for Testing AI Logic

RuleWeaver is the benchmark that finally tests whether AI truly follows complex rules in real-world scenarios. Most language model tests only check basic facts, but they miss the messy logic of actual life. This project solves that gap by weaving realistic situations from legal contracts and government reports, then asking models to navigate specific, tricky constraints. It does not just look for a correct answer; it verifies if the model used the right reasoning path to get there. This gives researchers a clear, honest way to measure if an AI is actually thinking or just guessing based on patterns.

🆔 @hackernewsgithubprojects

📦 sharkspicy-nlp/beyond-factual-knowledge

Teaching AI to Follow Rules

Teach a language model to actually follow rules, not just recite facts. That is exactly what Beyond Factual Knowledge does. It tackles a huge blind spot: models often ignore multi-step logic even when they have the right data. The team built a new benchmark and a training method that forces the AI to retrieve and apply specific procedural rules at the exact right moment in its thinking. Instead of just guessing, it learns to check a rule library step-by-step to solve complex problems correctly. It is a fascinating look at how we might push machines past simple trivia toward genuine, structured reasoning.

🆔 @hackernewsgithubprojects

📦 anthropics/commerce-agents

Commerce Agents: How to Build Smarter Store AI

Commerce Agents is the open blueprint that finally gives your business a smart digital front desk and a helpful back office manager. It is a collection of ready-made instructions for Claude that teach it how to help customers browse products and how to help staff manage inventory without breaking anything. The coolest part is that it shows you exactly how to build these assistants for different industries like travel or tickets, all using the same safe, proven code. It handles the tricky stuff like remembering what customers like and making sure no accidental orders get placed.

📰 https://news.ycombinator.com/item?id=49541574

🆔 @hackernewsgithubprojects

📦 mooreneural/lacuna

Find Hidden Protein Pockets

Find hidden binding pockets in proteins that appear completely flat. Lacuna solves a major hurdle in drug design by looking beyond a single static image of a protein. Many targets are called undruggable because no pocket is visible in their resting state, but they actually open up briefly. This tool generates a group of different shapes, or conformers, and scans each one to find temporary cavities. It then clusters these fleeting spots into stable sites and ranks them by how well they might bind a drug. The best part is that it runs in seconds on a standard computer without needing a powerful graphics card.

📰 https://news.ycombinator.com/item?id=49543481

🆔 @hackernewsgithubprojects

📦 hyacinthe-primus/file_converter_pro

File Converter Pro: Your Files Stay Home

File Converter Pro turns your Windows computer into a private, offline hub that changes any document, image, audio, or video file into a different format without ever touching the internet. This matters because your personal and professional files never leave your machine, meaning no uploads, no cloud servers, and no data breaches. It handles everything locally, so a Word document can become a PDF or a video can switch formats instantly, all on your hardware. The app runs entirely offline, ensuring total privacy for sensitive documents. You get full control over your data without sacrificing convenience.

🆔 @hackernewsgithubprojects