📦 vercel-labs/vgpu

Run Neural Networks On Your GPU With vgpu

Run real neural networks directly on your graphics card without juggling complicated frameworks. vgpu, built by Vercel Labs, gives you a simple toolkit that handles everything from fancy visual shaders to actual machine learning computations. The coolest part is that you can use the same code whether you are building a website or running scripts on your server. It even includes a ready-made example where it classifies handwritten digits using your GPU, proving it is more than just a toy for pretty visuals. This solves the headache of bridging the gap between creative coding and data science.

🆔 @hackernewsgithubprojects

📦 lennney/stop-that-shit

Stop That Shit: Your AI Coding Agent's Scope Creep Blocker

Ever ask an AI coding assistant to review a code change and watch it suddenly start rewriting half your project or adding useless checksums? Stop That Shit is a clever new tool built specifically to put a hard stop on that kind of scope creep. It plugs directly into popular AI coding agents and acts like a strict security guard for your software projects. If the AI tries to make changes you never asked for, the tool intercepts the action and blocks it, forcing the agent to stick to the exact task you assigned.

📰 https://news.ycombinator.com/item?id=49492705

🆔 @hackernewsgithubprojects

📦 cvhub520/x-anylabeling-server

One Server, Every Vision Model

One server, every vision model. X-AnyLabeling-Server is a lightweight Python framework that turns your AI models into a production-ready API, specifically built to power the X-AnyLabeling desktop app. Instead of juggling separate deployments for object detection, text recognition, and video segmentation, it handles it all through a single, flexible backend. The real kicker is its pluggable architecture: you can drop in new models like Segment Anything 3 or YOLO variants without touching the core code, while it manages the heavy lifting of concurrency, logging, and security.

🆔 @hackernewsgithubprojects

📦 open-webui/oikb

Oikb: Your Universal Knowledge Sync Engine

Stop copying files by hand. Oikb automatically keeps your Open WebUI Knowledge Bases fresh by syncing content from anywhere, including GitHub, Confluence, Slack, or even your local hard drive. It uses smart checksums to ensure only new or changed files are uploaded, saving time and bandwidth. The standout feature is its massive library of forty-six connectors, meaning you can pull data from almost any tool your team already uses without writing custom code. Once set up, it runs quietly in the background or on a schedule, ensuring your AI assistant always has access to your latest information.

🆔 @hackernewsgithubprojects

📦 miaai-lab/glm-5.3-flash-exl3-2x-dgx-sparks

Running a Massive AI Model on Tiny Desktop Chips

GLM-5.3 Flash is the setup that makes a massive AI model run surprisingly fast on two tiny desktop computers. It solves the problem of needing huge, expensive server racks just to run complex coding and reasoning models. By using a clever compression trick and splitting the work between two small NVIDIA boxes, this recipe lets the software think at speeds comparable to much bigger systems. The coolest part is how it handles long conversations without forgetting details. It’s a perfect example of how smart engineering can shrink big AI into your home lab.

🆔 @hackernewsgithubprojects

📦 xevrion/breakscale

Watch Your System Design Break

Breakscale is the system design simulator that actually lets you break things to see why they fail. Instead of staring at static diagrams or memorizing rules of thumb, you drag a slider to increase traffic and watch real requests queue up and stall. It runs a serious simulation where every number is measured, not guessed, so when your latency spikes or a retry storm crashes your database, you can see exactly how it happens. It’s the perfect way to build real intuition for distributed systems without needing a production environment. Stop guessing about system behavior and start experimenting with the code that actually breaks down under pressure.

📰 https://news.ycombinator.com/item?id=49488177

🆔 @hackernewsgithubprojects

📦 mirros-lab/code-as-world

Code-as-World: The AI That Codes Reality

Code as World is the AI framework that turns raw video into executable physics by writing the code behind the scenes. Instead of just guessing what happens next, an agent discovers and verifies a working simulation of the physical world, giving models a true grasp of reality. This repo makes it easy to run their cutting edge vision models locally, proving that teaching machines to code the world beats just memorizing pixels. It is a fascinating peek into how AI actually learns to understand objects and motion, not just recognize them. Check out this project to see how the next generation of physical reasoning is being built.

🆔 @hackernewsgithubprojects

📦 robbyant-research/zero-wam

Zero-WAM: Robots That Learn By Watching Humans

Zero-WAM is the robotic system that lets machines master brand-new tasks just by watching a short video of a human doing the job. This project tackles the hard problem of robot generalization, where previous systems struggled to perform tasks they had never physically practiced during training. Instead of needing thousands of hours of specific robot data for every single new activity, Zero-WAM uses human demonstration videos as visual instructions. The system watches how a person handles an object and translates those movements into actionable robot commands instantly. This allows the robot to figure out how to open a microwave or stack blocks without any prior experience or code changes.

🆔 @hackernewsgithubprojects

📦 amap-cvlab/abot-recon

ABot-Recon Rebuilds Long Videos Without Forgetting

ABot-Recon rebuilds long video streams into a 3D map by looking at only twelve frames at a time, keeping memory usage constant no matter how long the clip is. It predicts geometry and camera movement locally, then stitches those small pieces together to form a global reconstruction without needing to remember the entire past. This approach solves the problem of error drifting over long sequences while staying fast and efficient on standard hardware. It’s a clever way to handle long-term spatial awareness without heavy computational overhead, making it a strong tool for anyone working with extended video data.

🆔 @hackernewsgithubprojects

📦 huggingface/pwc-cli

pwc-cli: The Research Engine for Your Coding Agent

Your coding agent can now read the entire Papers with Code catalog without a single API key or account setup. Hugging Face released pwc-cli, a lightweight command line tool that lets AI assistants and humans query the largest open catalog of machine learning research. It isn't just a search bar; it gives your agent the ability to read paper abstracts, check for official code repositories, and browse benchmark leaderboards directly from the terminal. The real magic happens when you install the included skill, which teaches your agent exactly how to use these commands to ground its responses in actual scientific literature rather than hallucinated facts.

📰 https://news.ycombinator.com/item?id=49199801

🆔 @hackernewsgithubprojects

📦 mlightcad/awesome-cad

awesome-cad

Explore the full spectrum of open source computer aided design in this meticulously curated collection. It maps out the entire ecosystem, from classic desktop drafting tools to bleeding edge AI systems that turn simple text prompts into precise 3D models. The real star here is how it organizes the massive, often fragmented landscape of CAD kernels, parsers, and web based viewers into one clear roadmap. Whether you are scripting parts in Python or building browser based geometry engines, this list saves you from endless searching.

📰 https://news.ycombinator.com/item?id=49313259

🆔 @hackernewsgithubprojects

📦 kentucky-ai/opentakeoff

OpenTakeoff: The AI-Ready Construction Estimator

OpenTakeoff is the first open-source construction measurement engine that lets an AI agent and a human use the exact same tools to count materials on a building plan. It works entirely in your browser without uploading your files, and it features a one-click button that automatically detects room boundaries and calculates the square footage you need for bids. What makes it truly interesting is that it records exactly how every number was found, creating a clear audit trail for your construction projects. You can even let a computer program drive the whole process through a standard AI interface.

📰 https://news.ycombinator.com/item?id=49113911

🆔 @hackernewsgithubprojects

📦 token-gremlin/natural-disasters

Natural Disasters

Create cinematic storms with zero external assets. This browser project, called Natural Disasters, generates every wave, cloud, and hurricane pixel in real-time using pure math on your graphics card. There are no downloaded textures, 3D models, or sound files; the entire ocean and weather system is built from scratch at runtime. You can fly through a hurricane or trigger a tsunami with a single key press, and the water physically reacts to the disaster. Because everything is procedural, it runs smoothly in a standard web tab without heavy downloads. It is a perfect example of how far modern web graphics have come.

🆔 @hackernewsgithubprojects

📦 wtdcode/vllm-backport

vllm-backport: Run Frontier AI on Old GPUs

vllm-backport makes a 48-gigabyte A6000 card run frontier models like DeepSeek V4 Flash, which is genuinely wild. While the industry chases new hardware, this project takes the popular vllm engine and tweaks it specifically for older Ampere GPUs, like the 3090 and A100. The main trick? It uses a technique called full CUDA graph capture to squeeze a 47 percent speed boost out of your existing setup. It is not about buying new cards; it is about making your current hardware significantly faster and cheaper.

🆔 @hackernewsgithubprojects

📦 etiennepasteur/jean-claude

Jean Claude: The Proxy That Freezes API Settings

Jean Claude lets you freeze the exact settings a tool pulls from a remote server, stopping it from changing behind the scenes. It acts as a local traffic interceptor that answers specific requests with files you control, while everything else passes through untouched. This is perfect for locking down a configuration that shifts between sessions or for testing how a tool reacts to fake data without touching the real API. You simply define a rule in a config file and run your command through the proxy, keeping total control over what the software sees.

🆔 @hackernewsgithubprojects

📦 mlo-lab/levjepa

LeVJEPA: Video AI that thinks in real-time

LeVJEPA is the first video encoder that completely ditches the complex tricks usually required to stop AI models from forgetting what they learned. By using a clever mathematical constraint called SIGReg, this project trains a single, simple brain to watch videos without needing extra machinery to keep it from collapsing. The coolest part is that because the training is so efficient, the model can learn to understand the flow of time directly. This means it doesn't just see frames; it builds a memory of what happened before, allowing it to process new video in real-time without having to re-watch the whole clip every time.

🆔 @hackernewsgithubprojects

📦 lumberjacque/loadersz-core

150 Ways to Show Your App Is Thinking

Loadersz is the canvas animation library that finally makes waiting feel like watching something happen. Instead of a generic spinning circle, it offers one hundred and fifty distinct motion states that visualize specific actions, like weaving, solving, or breathing. Each state is a unique, hand-tuned visual that tells users exactly what the app is doing in the background. It runs as a tiny, dependency-free web component, meaning it fits into any modern framework without bloating your bundle. You can switch states at runtime or lock in a single animation for maximum performance. It is the rare tool that turns a mundane loading screen into a moment of clarity.

🆔 @hackernewsgithubprojects

📦 davidcarliez/trustmebro

TrustMeBro: Testing AI Safety with Fake Data

TrustMeBro is the red-team tool that lets you test if coding agents can be tricked by fake data. It works by standing in the way of the commands your AI uses, like checking internet addresses, and swapping in made-up results to see how the AI reacts. If the AI stops when it sees bad data, it’s safe. If it pushes forward, you’ve found a weakness. This project is about finding those cracks before bad actors do, making sure your AI tools are actually reliable and safe to use.

🆔 @hackernewsgithubprojects

📦 0xsero/local-ai-registry

The Local AI Registry

One machine can run an AI model, but knowing exactly which hardware makes it fast enough is usually guesswork. Local AI Registry fixes that by treating your computer specs like a strict filter for real software. It connects specific model files to your actual graphics card or Apple chip, then shows you verified launch settings and speed benchmarks. If the data says it works, it gives you the exact recipe to start it up. If the evidence is thin, it flags the setup as untested so you know exactly what you are risking. This removes the frustrating trial and error of local computing.

🆔 @hackernewsgithubprojects

📦 badtheorylabs/prism

Prism Makes Small AI Models Act Like Giants

Prism takes a small, local AI model and gives it the mental tools of a massive frontier model without needing the cloud. It works by moving intelligence from inside the model's weights to an external harness that handles context, planning, and verification. The surprising part is that it uses the compiler itself as the judge. By running code through tests and linters, it verifies the output objectively instead of letting the model guess if it is right. This turns a simple local app into a reliable coding assistant that checks its own work. You get frontier-level behavior on your own hardware.

🆔 @hackernewsgithubprojects

📦 iamlukethedev/hermes3d

Hermes3D: The 3D Office for Your AI Agents

Hermes3D is the open-source 3D engine that turns your scattered AI agents into a visible, walkable virtual office. Most developer tools hide behind dry dashboards and endless logs, but this project treats your AI workforce like a real team you can actually see working. You don’t just read about their progress; you watch them move around a retro-style 3D space, reviewing code, holding standups, and shipping tasks in real time. It’s like giving your AI a physical headquarters where you can monitor their activity and interact with them through a shared environment rather than a command line.

🆔 @hackernewsgithubprojects