📦 texel-org/windfoil

Windfoil: The AI-Born Rendering Trick

A single AI session produced Windfoil, a new algorithm that draws crisp 2D shapes with perfect anti-aliasing. It runs on your graphics card to calculate the exact amount of ink covering each pixel. Traditional methods often guess at edges, causing blurry or jagged lines on complex drawings. This new approach uses a winding integral to handle overlapping strokes and curves without extra cleanup steps. It solves a long-standing problem where thin lines or overlapping shapes would look messy or require heavy computer preprocessing. By computing coverage directly, it offers higher quality for things like generative art and detailed typography.

📰 https://news.ycombinator.com/item?id=49473043

🆔 @hackernewsgithubprojects

📦 miaai-lab/glm-5.3-flash-nvfp4-dual-dgx-spark

Run a Massive AI Model on Two Small Chips

Spin up a massive, multimodal AI model on just two small computer chips. This project gets a huge language model running on a pair of compact devices by splitting the workload across them using a smart networking setup. The biggest surprise is that it handles images and video while staying fast enough to be useful, proving you do not need a massive server rack to run advanced AI. It handles the complex wiring, memory management, and software setup automatically, so you can focus on using the tool. This is a clever way to build a powerful local AI helper at home without spending a fortune on hardware.

🆔 @hackernewsgithubprojects

📦 1038lab/comfyui-flashvsr

Real-time Video Upscaling in ComfyUI

ComfyUI-FlashVSR is the custom node that brings real-time video super-resolution directly into your ComfyUI workflow. It solves the problem of blurry, low-resolution video by using a diffusion model to upscale footage to 2x or 4x resolution without losing quality. The standout feature is its speed; it processes frames fast enough for streaming applications, thanks to optimized attention mechanisms and flexible model versions that fit different GPU memory limits. You can easily toggle between high-quality and fast-processing modes, and it even handles audio passthrough so your video stays in sync.

🆔 @hackernewsgithubprojects

📦 paoloanzn/pi-black

Pi Black: Use Your Claude Subscription

Pi Black lets you connect a third-party coding agent called Pi directly to your existing Claude Max or Pro subscription, completely bypassing the need for separate API keys. The project works by intercepting network requests and disguising them to look exactly like the official Claude Code application, ensuring the billing systems treat your new tool as a standard client. It handles the complex technical formatting required to pass authentication checks, including specific header calculations and identity verification. This is a clever workaround for developers who want to use their current flat-rate subscription across multiple different tools without paying extra for individual API usage.

📰 https://news.ycombinator.com/item?id=49473333

🆔 @hackernewsgithubprojects

📦 mirumo0u0/comfyui-cosmos-reference

Cosmos Gets Image Memory

The Cosmos model finally learned to look at your pictures. This new tool lets you feed a reference image directly into the latent space, essentially gluing it to your prompt so the AI can’t ignore it. It’s a brute force move that trades speed for raw visual fidelity, making sure every detail of your source material survives the generation process. While it runs a bit slower, it captures the essence of your input like a pro photographer would. If you need your AI images to actually match the source material, this is your cheat code for consistency.

🆔 @hackernewsgithubprojects

📦 tellahq/opensession

The AI That Lives on Your Computer

Open Session lets you run powerful AI coding agents directly on your own computer, giving you full control over your code. Unlike cloud tools that lock your work in the dark, this self-hosted server connects your Slack, Linear, and GitHub directly to an AI that actually writes and tests code in real git worktrees on your machine. It supports multiple AI models and lets you customize everything from the agent's persona to your specific repositories without forking the entire project. You get a secure, private environment where your data never leaves your infrastructure, and you can even use it from your phone via a simple web app.

📰 https://news.ycombinator.com/item?id=49473549

🆔 @hackernewsgithubprojects

📦 sapientinc/praxist

Praxist: The Autonomous Research Engine

Praxist is an autonomous system that runs a full scientific research loop without human intervention. It takes a project that already works and has a measurable goal, then coordinates parallel AI agents to explore different solutions. These agents test ideas, gather evidence, and use that data to guide the next round of experiments. This turns a static codebase into an active research process that can discover improvements you might not find manually. The system ensures every result is backed by verifiable evidence, making the entire process transparent and auditable. It is a powerful tool for developers who want their computers to do the heavy lifting of experimentation.

📰 https://news.ycombinator.com/item?id=49473502

🆔 @hackernewsgithubprojects

📦 db-aeon/joshu-oss

Joshu: Your AI Executive Assistant Workspace

Joshu builds a shared digital workspace where you and an AI agent operate the same software together, rather than the AI just typing on a screen. You run a private, self-hosted desktop on your own server that includes tools for email, a whiteboard, and a sandboxed browser. The agent can open files, search your notes by meaning, and handle tasks while you stay in the loop for sensitive actions. It’s a powerful way to create a persistent digital twin that actually works for you, not just against you.

🆔 @hackernewsgithubprojects

📦 naw103/foremerge

Foremerge: Catching AI Coding Agent Conflicts Before They Happen

Foremerge lets you catch intent conflicts before they turn into code disasters. Imagine two AI agents working on the same project in separate copies of your code. Git only sees text, so if one agent replaces a payment function and the other adds features to that same function, Git sees no overlap and merges them blindly. You end up with broken code that looks perfectly fine on the surface. Foremerge fixes this by having agents announce their plans first. They write down what they are about to change on a shared digital whiteboard stored inside your Git folder.

🆔 @hackernewsgithubprojects

📦 gabberflast/academic-pptx-skill

Make Claude Build Serious Academic Slides

Generate professional academic slides that actually argue a point, not just display data. This tool stops AI from creating flashy, empty decks and forces it to follow strict scientific standards instead. Every slide gets a clear title that states the main finding. The presentation flows like a logical story, with one chart per slide and proper citations. It removes unnecessary decorations so the science stands out. This is perfect for conference talks, thesis defenses, or grant proposals where you need to convince a skeptical audience with solid evidence, not just pretty pictures. Say goodbye to generic AI slides and get a deck that tells the real story.

🆔 @hackernewsgithubprojects

📦 tonyd2wild/glm-5.3-flash-nvfp4-2x-dgx-spark

GLM 5.3 Flash on DGX Spark

Run a massive, 320 billion parameter AI model on just two small desktop computers by fixing seven critical software bugs that had previously blocked everyone else. This repository provides the first known working setup for the GLM 5.3 Flash model on NVIDIA DGX Spark hardware, allowing it to process extremely long documents and generate text at high speeds. The creator solved deep technical issues in the underlying software stack to make this fast, energy-efficient hardware actually usable for such a large model. By sharing the exact fixes, the build files, and the performance tests, this project saves other developers weeks of trial and error.

🆔 @hackernewsgithubprojects

📦 tonyd2wild/glm-5.3-flash-nvfp4-dflash2-2x-dgx-spark

GLM-5.3 Flash on Spark: The First Working Setup

The first working deployment of the massive GLM-5.3 Flash model on two NVIDIA DGX Spark machines is finally here. This repository is not just code; it is a complete recipe for getting this thirty-two-billion-parameter model running on hardware that was previously impossible for it. The creator fixed seven critical bugs in the software stack, patching the engine to handle the unique memory layout of the new chips. The result is a stable server that can handle over two hundred thousand tokens of context, proving that even the most complex AI models can run on this compact, efficient hardware.

🆔 @hackernewsgithubprojects

📦 minnesotanlp/meta-n

Meta N: The AI That Codes Its Own Upgrades

Meta N is the research framework that makes AI agents upgrade their own code. Most systems just tweak final answers, but this project digs deeper. It uses a universal engine to read execution logs and write new helper functions directly into the agent's stack. Think of it as an AI that looks at its own mistakes and patches its own brain. This is not just theory; it is a working prototype that tests itself across nine major benchmarks, from legal reasoning to terminal commands. The result is a stack of self-improving layers that get better with every run. The takeaway?

📰 https://news.ycombinator.com/item?id=49458576

🆔 @hackernewsgithubprojects

📦 karuvanan/minimax-h3-director-cut-studio

MiniMax H3 Director Cut Studio

MiniMax H3 Director Cut Studio is the local video editor that treats AI generation like a professional film cut. Instead of just prompting a single clip, it gives you a Premiere Pro style timeline where you can plan shots, dialogue, and sound effects before rendering. The standout feature is its long-form assembly: it automatically breaks down your project into segments, renders them, and stitches them back together using the previous clip’s final frames to keep motion smooth. This solves the biggest pain point in AI video, which is consistency across long durations. You get a real production workflow with AI assistance, not just random clips.

🆔 @hackernewsgithubprojects

📦 perceptron-ai-inc/isaac

Isaac: The Robot Brain

Give a robot a brain that actually understands the world, and Isaac does exactly that. This open source model from Perceptron is a massive system that watches video and text to control physical robots, turning raw visual data into precise movements. The truly wild part is how it learns: it consumes one million hours of general video to figure out physics and space, which means it needs only twenty-eight hours of real robot practice to perform tasks that previously required thousands.

🆔 @hackernewsgithubprojects

📦 solis-team/xrepotest

XRepoTest: The Ultimate Test for AI Code

XRepoTest stops accepting AI promises and starts demanding proof by forcing language models to write unit tests for real code in five different programming languages. The project doesn't just look at the generated text; it actually compiles and runs those tests inside secure, isolated environments to see if they truly work. This setup solves the major problem of guessing if AI code is useful by measuring actual performance instead of just vibes. It checks if the code compiles, if it runs without errors, and how much of the original program it actually covers.

🆔 @hackernewsgithubprojects

📦 urbanground/urbanground

UrbanGround: Testing AI in Real Cities

UrbanGround is the realistic city simulator that actually tests if AI can walk down a real street. Most AI agents play in video games, but UrbanGround takes the entire three-dimensional layout of Hong Kong and turns it into a playable world where a language model has to navigate using only what it sees. You can literally watch the AI figure out how to climb stairs or find a coffee shop, and the project includes seven hundred specific challenges to see if it truly understands the space or just gets lucky.

🆔 @hackernewsgithubprojects

📦 irajisamurai/2d-motion-interface

2D Motion Interface: Captioning Video Without 3D

2D Motion Interface is the lightweight tool that lets AI describe human movement from flat video without needing expensive 3D tracking. Most motion models require deep, volumetric data that is hard to capture in the real world, but this project solves that by training a small bridge module. It teaches a standard 2D pose estimator to speak the same language as a 3D model, allowing it to generate accurate text captions from a simple phone camera. The surprising part is that this flat, 2D input actually outperforms traditional 3D estimates on real-world video, proving you can get high-quality results with far less computational power.

🆔 @hackernewsgithubprojects

📦 rarf/hermes-quota-plugin

Hermes Quota: Stop Guessing Your AI Limits

Hermes Quota is the desktop plugin that actually shows you how much AI usage you have left. It solves the mystery of hidden rate limits by pulling live data from providers like Anthropic and OpenAI directly into your sidebar and status bar. You get a clear progress bar for every account, so you know exactly when your credits reset without opening ten different tabs. The best part is it works quietly in the background, keeping your workflow smooth while it tracks your usage. No more guessing when you run out of tokens. It just works, giving you total control over your AI budget in one clean view.

🆔 @hackernewsgithubprojects

📦 links-ads/geoid-flood

GEOID-Flood: The Global Flood Dataset

GEOID-Flood is the satellite dataset that finally maps real-world floods across the globe with pinpoint accuracy. It pulls from two hundred nineteen actual disaster events spanning sixty five countries, turning them into thousands of high resolution images. The cool part is that it doesn't just show where the water is. It pairs satellite photos with terrain maps to teach computers the difference between a river that is always there and a river that has burst its banks. This gives researchers a huge, clean library to build systems that can spot dangerous flooding before it gets worse.

🆔 @hackernewsgithubprojects

📦 eximius-labs/fusion-embedding

Fusion Embedding: One Space for Seven Senses

A single model can now match a dog howling to the exact photo of that dog, without ever seeing a pair of audio and image during training. Fusion Embedding is an open-weight system that maps text, images, video, audio, thermal, motion, and touch into one shared vector space. It works by freezing a powerful base model and adding small, pluggable sensor heads, which means you can add new types of data without retraining the entire system. This is incredibly useful for developers who need to search across different types of media simultaneously, such as finding a video clip by describing its sound.

🆔 @hackernewsgithubprojects