📦 rauchg/gdp-ts

GDP-TS: Stop Authorization Bugs

GDP-TS is the TypeScript tool that makes authorization bugs impossible at compile time. It solves the classic problem where a developer forgets a check or mixes up user IDs, causing a security leak. Instead of just checking permissions, you pass a special ghost proof to sensitive functions. If the proof is missing or wrong, the code simply will not compile. It costs almost nothing at runtime and fits into any project. The type checker does the heavy lifting, so you catch mistakes before they ever reach production. It turns a risky manual check into a guaranteed safety net.

🆔 @hackernewsgithubprojects

📦 christianmat/jev-pokemon

Jev Plays Pokémon Red

Jev, the AI decision model from TypeSafe, is playing the entire main story of Pokémon Red with zero cheat codes or scripts. It runs inside a real emulator, reading the console's memory to see the world and then picking the very next button press itself. There is no hidden game knowledge here; it does not know where items are hidden or how to beat trainers without training. It has to figure it all out by walking, battling, and talking, just like you would. The best part is that it costs about a dollar to watch it play for a whole day.

🆔 @hackernewsgithubprojects

📦 andududu/jeview

Jeview: Watch Your Code Think

Jeview lets you watch every decision your code makes in real time. It runs on your own computer and sits between your app and TypeSafe, capturing each request and answer into a local database. The coolest part is the live map: it draws a growing tree of your logic as it happens, so you can actually see how one answer triggers the next. No cloud, no accounts, just a simple browser window showing your data flow. It’s a rare peek inside the black box, perfect for developers who want to understand exactly what their system is deciding and why. Give it a spin to finally see your logic move.

🆔 @hackernewsgithubprojects

📦 shootthesound/comfyui-fizgig-h3-still

comfyui-fizgig-h3-still

You can now extract stunning, crisp high-resolution still images directly from the MiniMax H3 video model without rendering heavy multi-frame clips or dealing with ugly, banded artifacts. A clever custom node extension called comfyui-fizgig-h3-still introduces two dedicated nodes that generate a true one-frame latent and decode it cleanly at virtually any size. By bypassing the default video decoding process, it preserves incredible skin detail and texture while running faster than standard workarounds. This makes it incredibly easy to create ultra-detailed image generation and editing workflows using your existing setup. Grab these nodes to start exporting gorgeous single-frame masterpieces today.

🆔 @hackernewsgithubprojects

📦 akkie76/code-review-skills

The AI Code Review That Doesn't Hallucinate Bugs

Code Review Skills is the instruction set that makes A.I. code reviewers actually stick to the facts. Instead of letting your coding agent guess or invent fake bugs to sound helpful, this tool forces it to provide concrete evidence for every single issue it finds. It completely ignores vague style complaints and only flags real problems with a clear impact. Think of it as a strict quality control layer for your A.I. assistants. If you are using tools like Codex or Claude Code to write software, this is the missing piece that keeps your reviews honest and useful. No more guessing, just verified facts.

🆔 @hackernewsgithubprojects

📦 psrben/visionhope

VisionHOPE: The AI That Learns While Looking

A new vision model called VisionHOPE doesn't just look at images; it learns to look at them in real time. Standard neural networks have fixed rules for processing data, but this one changes its own internal memory and learning rules as it scans a picture. This is called self-referential nested learning. Instead of using the same static filters for every pixel, the model adapts its entire processing strategy on the fly, based on what it has already seen in that specific image. This allows it to handle complex visual tasks like object detection and segmentation with a unique, dynamic approach.

🆔 @hackernewsgithubprojects

📦 lateos-ai/reflex

Reflex: The Cold-Start Inference Engine

Reflex is a Rust inference engine that delivers its first result in under half a second from a completely cold start. Unlike traditional servers that sit idle waiting for work, this tool runs one single task, finishes instantly, and shuts down to stop the meter. It precompiles its graphics math at build time, so there is no delay when the computer first wakes up. This makes it perfect for serverless platforms that charge by the second, where speed is the only thing that matters. It beats heavy competitors like vLLM by over three hundred percent on startup time. You get fast, single-shot answers without paying for waiting rooms.

📰 https://news.ycombinator.com/item?id=49852130

🆔 @hackernewsgithubprojects

📦 leaplabthu/simple-wam

Simple-WAM: The Robot That Dreams Before It Moves

Simple-WAM is a robot learning model that actually predicts the future video of what it is about to do before it moves a single joint. That sounds like sci-fi, but it is exactly how this system learns to generalize to new tasks it has never seen. Instead of just memorizing button presses, the model builds a mental simulation of the outcome, letting it adapt on the fly. This research code releases the exact training scripts and evaluation tools used in the paper, so you can see exactly how these world action models work under the hood.

🆔 @hackernewsgithubprojects

📦 storytold/filmcraft

FilmCraft: The AI-Ready Video Editor

An artificial intelligence can now cut, color grade, and mix your video exactly as skillfully as a professional editor, all without touching a mouse. FilmCraft is a complete, open-source video editor written entirely in Rust that runs natively on your desktop or in the browser. It recreates the familiar interface of Adobe Premiere Pro, but underneath, it is completely new, with its own codecs and a system where every single action is a command. This means any AI agent can drive the software to edit footage, adjust colors, or export files with the same precision a human has.

🆔 @hackernewsgithubprojects

📦 elripalda/porpoise-dolphin-emulator-for-ps5

Play GameCube on Your PS5?

A PlayStation Five can now run your GameCube and Wii discs natively, thanks to Porpoise Dolphin Emulator for PS5. This homebrew app ports the famous Dolphin emulator directly to the console, meaning no PC or phone required. The coolest part? It turns your DualSense controller into a fully functional Wii Remote, complete with motion tracking and pointer precision. It even adds a cover flow library with box art, memory card management, and original menu music. If you want to play those old favorites on your main TV without the clutter of extra hardware, this is how you do it. Just grab your backups and get ready to play.

🆔 @hackernewsgithubprojects

📦 djeada/standard-of-iron

Standard of Iron: RTS Meets Hack and Slash

Standard of Iron lets you command massive Roman armies while simultaneously hacking through enemies as a frontline commander. This open source strategy game drops you into a twisted Second Punic War, mixing grand scale tactics with visceral close quarter combat. The real surprise is the mode switch. You can zoom out to manage supply lines and formations, then instantly jump into direct control to fight with your own hands. It’s a clever twist on the genre that keeps you engaged at every level. If you love strategy games that don’t let you hide behind the curtain, this one is a must see.

🆔 @hackernewsgithubprojects

📦 sau412/esp32_cyd_pda

The ESP32 CYD PDA

The ESP32 CYD PDA is the tiny firmware that turns a cheap yellow display into a full handheld computer. It gives you a classic Palm OS feel without any extra hardware. You get notes, games, and a terminal on a single chip. The coolest part is the built-in oscilloscope for hacking with signals. It runs on Arduino code you can tweak. No soldering required. Just flash it and go. It is a neat little pocket lab. Perfect for tinkerers who want more than a clock. It proves small boards can do big things. Check it out if you love retro tech.

🆔 @hackernewsgithubprojects

📦 adadoom3/turboada

TurboAda: The Fast Ada Compiler

Compile Ada code at lightning speed with TurboAda. This new compiler uses LLVM to generate code that runs dramatically faster than the standard GNAT compiler. In benchmarks, it handles exceptions up to one hundred sixty-three times faster and memory allocation nearly four times faster. It fully supports the latest Ada standards, including tasking and generics, while maintaining complete conformance with all official tests. The tool is self-contained and works across Linux, macOS, and Windows. It even acts as its own language server for better editor support. If you write Ada code, this is the performance boost you have been waiting for.

🆔 @hackernewsgithubprojects

📦 ciprianveg/gb10-vllm

GB10-vllm: Massive AI Clusters on Your Desk

GB10-vllm is the open-source blueprint that turns a single NVIDIA DGX Spark box into a high-speed engine for running some of the biggest language models available today. Most people assume these massive AI systems require giant server farms with expensive cooling systems and endless racks of hardware, but this project proves that is no longer true. It provides the exact setup needed to run powerful models like Kimi K3 and GLM directly on compact desktop hardware. This is fascinating because it democratizes access to top-tier AI without the massive infrastructure overhead.

🆔 @hackernewsgithubprojects

📦 animatek/g1-emu

G1-Emu: Run the Nord Modular G1 on Your Computer

Turn your computer into a full-sized Nord Modular G1 synthesizer. This project emulates the original hardware, meaning the actual operating system and digital signal processors run inside your machine just like the real rack unit. It is not just a software synth pretending to sound like it; it is the real thing running on emulated chips. The coolest part is that you can connect any compatible editor to it, so you get that authentic modular workflow without the massive hardware footprint. It is a pre-alpha community effort, but it already plays clean audio in real time.

🆔 @hackernewsgithubprojects

📦 nv-tlabs/pixelumm

PixelUMM: One Model for Pixels and Words

PixelUMM is the image and video engine that skips the usual translation step. Instead of using separate parts to look at pictures or make new ones, this single model works directly with raw pixels. It uses one smart network to both understand what it sees and create brand-new images and videos from your text prompts. This means no complicated conversion between formats, just direct interaction with the visual data itself. It is fascinating because it simplifies how computers handle visual media, making the process cleaner and more direct. You get one tool that can look at a clip and then generate a whole new scene based on what it learned.

🆔 @hackernewsgithubprojects

📦 deltarooo/er-mario

er-mario

You can now play through the entire world of Elden Ring with the exact physics, movement, and controls of Super Mario 64. The er-mario project integrates the legendary platformer's original engine directly into the dark fantasy action game using live collision data. This means you can run, triple jump, wall kick, long jump, and ground pound across the Lands Between, using Mario's classic moves to traverse massive cliffs and engage in combat. You can even punch enemies, swing staggered bosses by the tail, and throw them like Bob-ombs while managing a classic retro health bar.

🆔 @hackernewsgithubprojects

📦 nealbridges/vulnhunter

VulnHunter: Security Scanning That Actually Works

VulnHunter is the agentic AI security scanner that actually proves vulnerabilities are exploitable before telling you about them. Instead of just flagging suspicious code patterns like traditional tools, it acts like a hacker, walking through the code to find real attack paths. The most surprising part? It doesn't just find bugs; it writes and runs actual tests to prove the flaw works. If it can't prove it, it throws it out. This means you get verified, actionable fixes with real evidence, not a mountain of false alarms. It's a game-changer for anyone who wants to know exactly what's broken, and how to fix it, without the guesswork.

🆔 @hackernewsgithubprojects

📦 academiasd/academiasd_loralab-trainerstudio

AcademiaSD LoRAlab Trainer Studio: Train AI Models on Your Laptop

AcademiaSD LoRAlab Trainer Studio is the unified training hub that lets you teach custom styles and characters to the newest image and video models using just a mid-range graphics card. It solves the problem of heavy setup by bundling everything into one simple launcher that handles the complex math for you. The surprising part is that it runs these massive models on as little as four gigabytes of video memory by compressing them in a smart way called four-bit quantization. You get one web interface to manage datasets, watch live previews, and export your finished training files directly into ComfyUI.

🆔 @hackernewsgithubprojects

📦 uditakhourii/quicksilver

Quicksilver: Cut Your Claude Token Bill by 86%

Handing massive amounts of text to a fast, cheap model lets your main AI assistant skip the boring parts and focus only on what actually matters. Quicksilver works as a special skill for Claude Code, automatically routing bulk judgment tasks like filtering logs or finding specific files to a cheaper system called Jev. This tiny model scans hundreds of items in parallel, returning a tight shortlist of yes or no verdicts in about a second. By keeping the noise out of Claude's context, the tool cuts token usage by eighty-six percent and speeds up work by up to twenty times.

🆔 @hackernewsgithubprojects