📦 jianglongye/dexterous-astra

Dexterous Astra: A Hand That Spins a Pen

A robot hand can spin a pen and twist a Rubik's cube. This project, called dexterous Astra, shows how simulated robot hands master these dexterous tricks. It provides ready-to-run environments and trained policies. You can run a zero-shot script that simply tells the hand to spin the pen. Or you can load a trained policy to see complex moves. The goal is to study how hands interact with objects. It is a clean way to test manipulation skills. You don't need to build the physics engine yourself. The code is set up for easy experimentation. This is a great resource for anyone studying robot hands. It makes complex dexterity accessible.

🆔 @hackernewsgithubprojects

📦 blader/humanizer

Humanizer: Make AI Text Sound Human

Take your AI generated prose and make it sound like a real person actually wrote it. Humanizer strips away the robotic habits that make text feel artificial, like forced lists or dramatic pauses, without changing your facts. It spots twenty five specific tells where language models slip up. The coolest part? It matches your unique voice if you paste a sample of your own writing. No more sounding like a chatbot. Just clean, natural prose that is actually yours. Give it a try.

📰 https://news.ycombinator.com/item?id=47362115

🆔 @hackernewsgithubprojects

📦 mutonby/vibetube

VibeTube: Record, Edit, and Publish with AI

VibeTube is the Mac app that turns raw multicam recordings into finished, publish-ready videos using AI. It records your screen and webcam in perfect sync, then hands the footage to an AI agent that edits the final cut, adds subtitles, and writes the description. You never have to touch a timeline or open a browser again. The AI even publishes the video directly to YouTube and TikTok for you. It is the closest thing to a fully automated video production pipeline available right now. Stop spending hours on editing and let the machine do the heavy lifting.

🆔 @hackernewsgithubprojects

📦 andrezaiats/snes9x-z

Snes9x-Z: The SNES Emulator That Runs 60% Faster

Snes9x-Z, a tweaked SNES emulator, squeezes out over sixty percent more speed from your computer without changing a single pixel of the output. This is not a magic trick; it is a carefully built fork that takes the standard emulator and rearranges the internal work so the graphics engine uses two CPU cores instead of one. The result is a massive performance boost that still produces bit-for-bit identical results to the original. Why does this matter? Because those extra cycles let you run shader filters, record gameplay, or play demanding games on weak hardware like old handhelds without lagging.

🆔 @hackernewsgithubprojects

📦 steadfastasart/geoscience-skills

Geoscience Skills: The AI Tool For Earth Data

Geoscience Skills is the open-source toolkit that finally makes coding agents actually understand the messy reality of earth science data. Most AI tools stumble when they encounter specialized formats like seismic files or well logs, but this project fixes that by teaching them exactly how to handle thirty different scientific libraries. Instead of guessing, your coding assistant learns to properly read a seismic file or calculate rock properties using the right formulas. It turns generic chatbots into genuine geological assistants capable of real analysis. If you work with earth data, this is your shortcut to having a smart partner who actually gets the science.

🆔 @hackernewsgithubprojects

📦 shadcn-labs/pdfcn

Turn React Code into Polished PDFs with pdfcn

Turn your standard React components into polished, downloadable PDF documents with just a single command. This free, open-source tool from the shadcn team solves the common headache of making complex web layouts look professional on paper. It uses two distinct rendering engines, allowing you to generate clean invoices, reports, or data tables without wrestling with obscure formatting settings. Because it follows the same familiar workflow as other popular UI libraries, developers can copy, paste, and ship high-quality documents immediately. You own the code, meaning you can fully customize every detail to match your brand.

🆔 @hackernewsgithubprojects

📦 xingchen-agi/xing4.0-29b-a4b

Xing 4.0: The Open Source AI Trained on Chinese Chips

Xing 4.0 is the open source model that proves domestic hardware can rival global giants. It is a massive language model built by China Telecom, featuring a design that activates only a fraction of its brain to process huge documents efficiently. The standout feature is that it was trained entirely on Ascend processors, offering a real alternative to traditional silicon. This makes it a powerful tool for developers needing strong coding and reasoning skills without relying on imported chips. You get a high performance engine that runs natively on this new platform, ensuring stability for complex tasks.

🆔 @hackernewsgithubprojects

📦 alibaba-damo-academy/damo-radar

RADAR: AI that reads CT scans

Diagnose abdominal conditions directly from CT scans using RADAR, a generalist AI model. This project solves the problem of needing separate tools for different diagnoses by learning directly from over forty thousand clinical reports without manual labels. It identifies twenty specific conditions, like aortic aneurysms or kidney cysts, with expert-level accuracy. The standout feature is its ability to understand anatomy-aware image-text pairs, bridging the gap between visual data and clinical language. This open-source framework provides a scalable path for radiology AI, offering researchers and developers a robust foundation for building next-generation diagnostic tools that work across complex clinical tasks.

📰 https://news.ycombinator.com/item?id=49761981

🆔 @hackernewsgithubprojects

📦 standardagents/jevpilot

JevPilot: The AI Driving Demo

This browser-based driving simulator lets an AI drive a Tesla Model Y through a realistic 3D city by sampling dozens of potential steering paths every fraction of a second. Built with Three.js and powered by Jev, it visualizes the exact moment the algorithm chooses its route, painting the selected path in bright blue while ghosting the rejected options in cyan and amber. The coolest part is that you can flip a switch to see every possible lane change and collision prediction before the car commits to a move.

🆔 @hackernewsgithubprojects

📦 dartnative/dartnative

DartNative: Real Native Apps Written in Dart

DartNative is the mobile framework that writes your app in Dart while using the phone’s actual built-in buttons instead of drawing fake ones on a screen. Usually, cross-platform tools paint their own graphics, making apps feel slightly off or slow to touch. DartNative skips that middleman entirely. Your code talks directly to the operating system’s real interface parts on a single, fast track. This means your app looks exactly like it belongs on the device, moves with the system’s natural speed, and feels instantly responsive. It’s the same coding style you know from Flutter, but the result is a true native application without the extra rendering weight.

🆔 @hackernewsgithubprojects

📦 openbmb/simplememvla

SimpleMemVLA: Robots That Remember Like Humans

A real dual-arm robot can now look back at its own video feed to remember exactly where it picked up a block, a capability called SimpleMemVLA. Instead of using complex memory modules that might lose details, this system treats the robot's past footage as a native video. It feeds this timestamped history directly to the brain of the robot, allowing it to figure out what happened minutes ago. This simple approach lets the robot succeed at tricky tasks like covering blocks in a specific order, achieving high success rates in real-world tests. By keeping the raw video intact, the robot makes faster and more accurate decisions without confusing similar objects.

🆔 @hackernewsgithubprojects

📦 0xsero/qwen38-3090-sglang

Run Qwen 3.8 on a Single GPU

A single consumer graphics card can now run a massive multimodal AI model that understands both images and videos at speeds that rival expensive professional hardware. This repository provides a ready to use Docker container that packages the Qwen 3.8 twenty seven billion parameter model specifically for NVIDIA RTX 3090 graphics cards. It solves the complex problem of making this large model run efficiently on standard PC hardware by using specialized quantization and a unique decoding technique called DSpark. The container handles all the difficult software setup automatically, so you just need to run one simple command to start serving requests locally.

🆔 @hackernewsgithubprojects

📦 chetaslua/jevmeter

Jevmeter: The BS Meter for Any Video

Jevmeter is the video tool that turns every sentence into a live scorecard. It uses an AI model to analyze your footage and render a cool, sixteen-by-nine edit that flags evasive answers and hype in real time. The best part is that it costs just five cents to process an entire debate. You drop in a file, pick a preset like debate or sales pitch, and it handles the rest. It is the perfect way to turn any talk into a shareable highlight reel. Give it a spin and see who dodges the most questions.

🆔 @hackernewsgithubprojects

📦 tamaratran/fast-jev-compaction

How fast-jev-compaction saves your code context

Shrink your coding bot’s memory without losing the details that matter. Fast Jev Compaction fixes the annoying habit of AI tools forgetting file paths or exact errors just to fit more text into the window. Instead of summarizing your chat, it asks a second model to look at your entire history and decide exactly which old tool results to drop. Anything it keeps stays word for word, so you never lose a critical command. This plugin turns messy, bloated conversations into a clean, precise record. Your bot remembers the important stuff, not just a vague summary. It’s a simple swap that keeps your coding sessions sharp and accurate.

🆔 @hackernewsgithubprojects

📦 arahim3/cachebeat

Cachebeat: Stop Paying Full Price

Cachebeat is the tiny Claude Code skill that stops your conversation from paying full price just because you stepped away. It solves a sneaky billing trap where long AI sessions lose their memory discount after an hour of silence, causing the entire chat history to be re-processed at ten times the normal cost. The surprising part is that it uses a simple inactivity timer to send a few words of text, like saying "still here," which resets the clock and keeps the cache warm. This means your next message reads from the cheap cache instead of triggering an expensive rebuild, saving you real money or protecting your subscription quota.

🆔 @hackernewsgithubprojects

📦 magiccodingman/vllm-radiance

VLLM Radiance: The AMD GPU Inference Secret

This project runs large language models on AMD Radeon AI Pro R9700 graphics cards by compiling a custom version of the VLLM inference engine specifically for that hardware. Most standard AI servers are built for Nvidia chips, so running complex models on AMD cards is usually slow and unstable. This repository fixes that by using hand-written math routines that match the specific architecture of the R9700 GPU, allowing it to serve text and vision data at high speeds. It is an experimental but highly tuned setup that proves you can get serious performance out of consumer-grade AMD silicon.

🆔 @hackernewsgithubprojects

📦 scrypthunter/comfyui-yue2

ComfyUI Yue2: The Safe Song Maker

ComfyUI Yue2 is the music node that finally lets you generate full stereo songs without crashing your setup. It solves a nasty problem where installing new music tools would break your existing image generation software. The fork loads models safely in isolation, so you can turn text and lyrics into a complete track or extract a score from a reference song to create a cover. It fixes the dependencies that usually cause conflicts, meaning your whole system stays stable while you create music. If you want to make songs on your computer without losing your image workflows, this is the one to try.

🆔 @hackernewsgithubprojects

📦 theoleecj/semif

SemIf: AI Decisions Without the Fluff

SemIf is the open source tool that reads AI decisions directly without waiting for text to generate. Usually, when you ask an AI to make a simple choice, like routing a ticket or verifying a fact, it has to write out a full sentence before your software can actually understand the answer. It is slow and wasteful. SemIf skips that entirely. Instead of forcing the model to speak, it peeks directly at the raw numbers inside the model to see which option is most likely. This means you get the answer in about one second instead of five.

🆔 @hackernewsgithubprojects

📦 maksshandybo/perestruct

PereStruct: The Soviet Newspaper Solver

PereStruct is the document pipeline that finally makes sense of messy Soviet newspaper scans. It takes a chaotic page of text, uses computer vision to find where the articles start and stop, and then uses language models to clean up the errors. The coolest part? It figures out which scattered blocks belong to the same story, even if they are on different columns. It actually beats out massive vision-language models at this task. So if you have a pile of old, hard-to-read news clippings, this is your secret weapon for turning them into clean, readable text.

🆔 @hackernewsgithubprojects

📦 opendfm/image_edit_agent

Chat to edit photos like a pro

Fix your photos just by talking. A new research project called Image Edit Agent turns casual speech into precise photo adjustments, letting you tweak exposure or contrast without touching sliders. Instead of generating random images, it controls standard editing tools so you can see exactly what changed and fix mistakes easily. This makes professional retouching accessible to anyone, turning a complex task into a simple conversation. It is a practical step toward truly understandable and controllable AI image editing.

🆔 @hackernewsgithubprojects

📦 jarrodwatts/jev-trader

Jev Trader: AI Trading at 300 Milliseconds

Place a real trade on the blockchain every three hundred milliseconds. Jev Trader uses an AI model called Jev to watch the order book for the Monad and USDC pair. Every single block, the AI makes a split-second decision to buy or sell. It then sends a live order from a real wallet, landing in the same block it made the choice. The whole process, from reading the market to posting the trade, fits into a tight 300-millisecond window. A dashboard streams every decision, order, and fill live, so you can watch the AI trade in real time.

🆔 @hackernewsgithubprojects