298 subscribers
14K videos
14.6K links
Download Telegram
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ everruns/everruns

Everruns: The Engine for AI Agents That Never Forget

Run complex AI agents that actually finish their jobs, even when they take hours. Everruns solves the annoying problem where an agent crashes or loses its memory mid-task, starting everything over from scratch. It works like a durable checkpoint system, saving the agent's progress so it can pick up right where it left off without losing any context. This is built in Rust for serious reliability, letting you scale agents across different AI providers without the usual headaches. If you want your AI to handle long, multi-step tasks that actually stick around, this is the engine you need.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ luwudynamics/xgoduck_runtime_arduino

XGO Duck: A Tiny Robot That Walks

This tiny duck robot actually picks things up using real-time neural networks. It runs on an Arduino Uno Q, where a tiny chip handles the legs while a bigger chip runs the brain. The cool part? If the duck falls over, the software detects it and automatically triggers a recovery model to get it back on its feet. This project bridges the gap between hobbyist hardware and serious robotics. You get a web interface to control its walk, head, and even its mouth. It is a fantastic example of how efficient hardware can handle complex behavior without a supercomputer.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ velo4705/awesome-free-byok-models

Awesome Free BYOK Models: No Credit Card Needed

This list only includes AI model providers where you can grab an API key without ever touching a credit card. It solves the annoying problem of fake free trials that quietly expire or demand payment details upfront. Every entry is tested by real people to confirm it stays free and works instantly. You get a simple table showing exactly which models run, how much data they handle, and how fast they respond. No hidden costs, no surprise bills, just verified access. It cuts through the noise so you can find a reliable, zero-cost API that actually works for your projects.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ ichris4/wiicompiled_vr

Mario Kart Wii Runs Natively on PC

WiiCompiled VR is the native port that finally lets you race Mario Kart Wii on your PC without a single line of emulated code. Instead of faking a Wii console with an interpreter, the project statically recompiles the game’s original instructions into native C++ that your computer actually executes. The standout feature is experimental virtual reality support, where you sit directly in the cockpit. Your steering wheel turns in sync with your tracked controllers, creating a genuine sense of driving rather than just looking at a screen.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ bornholm/indecis

indecis

indecis lets you build tiny decision models that answer specific questions about text with calibrated probabilities instead of generating free-flowing replies. It takes a small pretrained encoder and fully fine-tunes it on your own labeled examples so it can classify topics, detect intent, or score urgency in milliseconds. The whole thing runs on a standard laptop without a GPU, using less memory than most browser tabs and training in just thirty minutes on twenty thousand examples. If you need fast, reliable, and private text classification, this is a surprisingly capable tool. It proves you don't need a massive AI infrastructure to get smart, structured answers from your data.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ bladedevoff/stuntd

stuntd: The AI Proxy That Learns Your App

stuntd acts as a smart middleman that watches how your app uses large language models to make specific decisions, like sorting tickets or checking if a command is safe. Instead of paying for every single question, it quietly records those answers and trains a tiny, lightweight model to replicate them locally on your own hardware. This means it can answer the same questions in about twenty milliseconds without ever sending your data to an outside server. If it is unsure, it safely hands the job back to the main provider, ensuring accuracy while cutting down on costs and latency.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ leonxlnx/verdant-forest

Verdant Forest

Verdant Forest is the browser-based woodland that renders thousands of unique trees in real time. This project builds a dense, temperate forest directly in your web browser using actual curved geometry rather than simple flat images. You can fly through the scene with your keyboard to explore thousands of individual grass blades, ferns, and wildflowers. The real surprise here is the scale. The engine places over seven thousand distinct trees and millions of small ground details, all moving and reacting to wind simultaneously. It solves the problem of digital nature feeling flat by creating a living, breathing environment that you can actually walk through.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ rhfeiyang/geb

GEB: The AI That Remembers Objects, Not Just Time

GEB is the long-video memory framework that finally stops confusing two identical mugs. Instead of just listing events in order, it builds a distinct biography for every single physical object it sees. This means the system knows exactly which red cup went into the dishwasher, even if another identical one stayed on the counter. By linking visual sightings across hours, it tracks the actual history of specific items rather than just generic descriptions. It’s a clever way to give machines real-world common sense about identity. If you work with video data, this is the kind of detail that actually matters.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ leonvanzyl/cubefarm

CubeFarm: The 3D Office for AI Developers

CubeFarm is the cartoon three-dimensional office that finally lets you watch your code get written. It turns a list of tasks into a living building where a team of artificial intelligence workers sits at desks, coding through your project issues. The most surprising part is that you can literally walk up to their monitors to see their real terminals running, or watch the quality assurance testers open your actual web app in a browser to verify the fixes. It is a fun, transparent way to manage automated software development without staring at a wall of text.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ gitcommit90/glm-5.3-one-spark

Run GLM 5.3 on One Spark

Run a massive language model on a single NVIDIA DGX Spark with GLM 5.3 One Spark. This recipe squeezes the GLM 5.3 Flash model onto one card by using aggressive compression and speculative decoding tricks. It handles huge context windows and serves requests through a standard API. The project publishes real benchmark numbers so you know exactly what speed to expect. It is a practical setup for developers who want local inference without a cluster. You get a working system with clear performance data. This is a solid reference for local AI hardware. It shows what is possible with smart optimization. Check it out if you are building local AI.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ joshuagodwin7929/purple-team-automation

Purple Team Automation: The AD Lab Test

This project automates a full Active Directory credential theft chain to test if your security alerts actually catch real attacks. Instead of running one isolated trick, it uses a framework to execute a realistic sequence of password spraying, ticket requests, and data syncing, all at once. The cool part is finding the blind spots. It turns out that while the system caught four out of five techniques, a very common poisoning attack slipped through completely because the underlying logging wasn't even turned on. It is a practical way to prove your detection rules work, not just assume they do.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ dime-online/qwen3.8-flash-dgx-ultrafast

Qwen3.8 Flash DGX UltraFast

Run a blazing fast AI model on a single, pocket-sized computer. This repository gives you the exact recipe to squeeze seventy-four tokens per second out of one tiny chip, which is wild for hardware that size. The magic trick is speculative decoding. A smaller, faster helper model guesses the next words, and the big model just checks the math. Since the guesser is correct most of the time, you get massive speed boosts without any loss in answer quality. You don't need a data center rack. You just need one card and this setup guide. It turns a small box into a local server that feels instant.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ musoyangrigor/scroll-video-website-skill

Scroll Video Website

Scroll Video Website is the AI skill that turns any video into a smooth, scroll-controlled website. Instead of playing a video, it breaks the clip into lightweight images so you can scrub through frames like a movie. You simply point your AI agent at the file, and it builds a responsive page where scrolling drives the animation. This is perfect for creating cinematic product reveals or landing pages that feel premium and interactive. The best part is that it optimizes the images to load fast while keeping the motion buttery smooth. It is a clever way to make your web projects feel dynamic without heavy coding.

πŸ“° https://news.ycombinator.com/item?id=49401459

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ santoshyadavdev/angular-devtools

Angular DevTools: The Inspector for Agents

Angular DevTools is the inspection tool that finally lets your AI coding agents see your app. Most debuggers show you code, but this tool actually looks at your running application to see how components, signals, and routes behave in real time. It acts as a bridge between your browser and your coding assistant, giving it full visibility into your app’s live state without you having to manually describe the issue. This is huge because it allows your AI helper to verify its own work directly against the running interface. It stops the guesswork entirely. Next time your code misbehaves, hand the reins to your agent.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ concile-dev/concile

Concile: Your Entire Backend in One Binary

Concile packs a database, real-time data syncing, user authentication, file storage, and scheduled background jobs into a single executable file you can host on any server. Instead of spending weeks wiring together APIs, managing complex cache layers, and writing thousands of lines of boilerplate glue code, you just write plain TypeScript functions and let the platform handle the rest. The most surprising part is how it keeps everything in sync without you ever touching a refresh button or writing a single line of socket code. When you update a piece of data on your server, every screen currently looking at that data updates instantly across the entire app.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ facebookresearch/swe-sweep

SWE-sweep: The Bug Hunter Benchmark

SWE-sweep gives AI agents a real codebase and lets them hunt down bugs on their own, with zero hints about where to look. It is a stress test for large language models, asking them to find and fix as many issues as possible in complex software. The interesting part is that the AI has to figure out what is broken without any guidance, just like a human developer debugging a tricky app. This project, built by Facebook Research, helps us understand how well AI can actually handle real-world coding problems instead of just simple exercises.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ gillespiejameson/personal-finance-tracker

Personal Finance Tracker

Drop a raw bank statement into Personal Finance Tracker and watch it automatically identify the bank, fix the dates, and merge the data into one clean searchable table. This local first tool keeps your entire financial history in a single file on your own computer, so there are no cloud servers or accounts to manage. It goes far beyond a simple ledger by spotting recurring bills and paychecks, then calculating a real safe to spend number alongside your total net worth. You also get cash flow forecasts and anomaly alerts that warn you when spending suddenly shifts out of your normal pattern.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ boxabirds/awesome-local-ai

awesome-local-ai: The One-Command Way to Run Real AI Locally

Get a powerful AI assistant running on your own hardware in just one command. This project cuts through the endless setup tutorials by giving you ready-made installers that automatically pick the best model for your specific machine. The coolest part is that they actually test these setups to see if the AI can build real software, so you know it works. No guessing, no broken configs. Just one script that finds your perfect fit and gets you coding with local intelligence immediately. It’s the shortcut to private, offline AI that actually respects your hardware. Skip the setup pain and start creating with a tool that’s ready to go.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ dddevid/spotlight

SpotLight: Spotify without the bloat

The official Spotify desktop app weighs in at over four hundred megabytes and often eats gigabytes of memory, but SpotLight proves you don't need a web browser to listen to your music. Built entirely in Rust, this ultra-lightweight client replaces the heavy Electron wrapper with native code that runs on just sixty megabytes of RAM. It connects directly to your Premium account to play audio, search tracks, and scroll through synced lyrics without the background bloat that drains your laptop battery. When you aren't interacting with the window, the interface completely goes to sleep, keeping CPU usage near zero.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ morrowmake/glm53-flash-cmp170hx-recipe

GLM 5.3 Flash Recipe

A 320 billion parameter model is now served locally on just four mid-range graphics cards, pushing nearly a thousand tokens per second for eight simultaneous users. This repository, known as the GLM 5.3 Flash recipe, solves the problem of running massive AI models on hardware that was never officially designed for them. It works by using a specialized engine that adapts to specific graphics chip architecture, allowing a single command to install and launch the entire system in an isolated container. The real surprise is that it achieves this speed without cutting corners on data precision or offloading memory to your disk.

πŸ†” @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
πŸ“¦ eliovp-bv/paiton-vllm-plugin

Paiton: Run AI on AMD Radeon

Paiton is the AMD GPU plugin that finally lets you run heavy AI models locally without a proprietary cloud. It works by compiling your specific AI model directly into optimized code for your AMD Radeon graphics card, which means faster speeds and better privacy for your data. You can run everything from text chats and coding assistants to image and video generators, all on your own hardware. Whether you are a developer testing a new model or a creator looking for local tools, this gives you total control over your AI stack.

πŸ†” @hackernewsgithubprojects