📦 soul-ailab/soulx-flashtalk

SoulX-FlashTalk: Real-Time Infinite Streaming of Audio-Driven Avatars

Generate endless, real-time video of a person talking directly from audio input. SoulX-FlashTalk solves the slow startup problem for large video models by launching in under a second. It streams video frames at thirty-two frames per second on high-end hardware, creating avatars that appear to talk forever without breaking. The key feature is its ability to correct itself as it generates, ensuring the face stays stable and lifelike throughout long conversations. This makes it perfect for live digital humans that need to feel immediate and present. You get a fast, continuous talking head without the usual lag, making it a standout tool for real-time interaction.

🆔 @hackernewsgithubprojects

📦 lyehe/build_gpu_colmap

COLMAP GPU Builds: No Install Hell

Building COLMAP for GPU acceleration usually turns into a nightmare of compiling dependencies and fighting drivers, but this project hands you ready-made binaries that just work. It packages the photo-to-3D tool with NVIDIA support so you can skip the build scripts entirely. On Windows, the files are huge because they include everything needed to run, meaning you don't have to install the CUDA toolkit separately. On Linux, they stay small by relying on your system libraries. You get fast reconstruction speeds without the usual setup headaches. It is the shortcut for anyone who wants results, not a computer science degree. Save yourself the pain and grab a pre-built package today.

🆔 @hackernewsgithubprojects

📦 nb-corp/gpt-with-taste

GPT with Taste: Better Judgment, Less Noise

GPT with Taste makes your AI coding assistant act with actual judgment instead of just guessing. It drops a special recipe into OpenAI Codex to stop the model from burying your results in defensive disclaimers and unnecessary process talk. The standout feature is how it handles getting-started guides. Instead of leading with environment checks and failure handling, it hands you a working example and tells you exactly what to click to see it change. It cuts the ceremony and gets you to the actual code faster. It is a simple fix for AI that talks too much and builds too little. Try it out and see the difference.

🆔 @hackernewsgithubprojects

📦 chenglou/handwired

Handwired: JavaScript as a Neural Net

Handwired turns regular JavaScript code into a neural net without any training. It converts standard logic like loops and branches into simple math that runs exactly like your original code. This lets you search a hundred thousand documents in milliseconds by using your graphics processor to parallelize tasks normally done one at a time. The result is a fully differentiable program that acts like a neural network but needs no data to learn from. This approach makes it possible to attach specific logic to larger machine learning models for end-to-end training. It’s a clever way to merge traditional programming with deep learning for high-speed, flexible applications.

🆔 @hackernewsgithubprojects

📦 crocswap/integer-mult-bounds

integer-mult-bounds

A crowd of independent researchers just managed to shave a tiny sliver off the theoretical speed limit of multiplying giant numbers. This project, called integer mult bounds, isn't about making your phone faster right now; it is about proving that big number math can be done with just a hair fewer steps than we thought possible. It solves a deep puzzle in computer science by building a massive, complex network of calculations that reuses parts of the work to save energy. This is pure math, not a quick fix, but it shows how a global community can collaborate to push the boundaries of what is computationally possible.

🆔 @hackernewsgithubprojects

📦 lucky-verma/mwb-linux

Control Windows from Linux with Mouse Without Borders

Move your mouse to the edge of your Linux screen, and the cursor instantly hops over to your Windows PC, taking your keyboard and clipboard along for the ride. This project, called Mouse Without Borders for Linux, solves the annoying problem of needing two separate sets of input devices when you run both operating systems. It acts like a magical bridge, letting you copy text on one machine and paste it on the other without touching a USB cable. The coolest part is that your Linux mouse stays perfectly still on the desktop while your digital avatar moves on Windows, preventing any accidental clicks.

📰 https://news.ycombinator.com/item?id=48823108

🆔 @hackernewsgithubprojects

📦 nutlope/gloss

Gloss: The Reader That Answers In The Margin

Gloss is the reading tool that finally lets you talk back to what you are reading. Instead of switching to a search bar, you simply highlight a sentence and type your question. The artificial intelligence answers right there in the margin, pointing to the specific paragraphs it used. It works on articles, books, and PDFs. The design feels like a physical notebook, with the text on the left and the AI notes on the right. It keeps your notes and highlights safe on your device, so you do not need an account to start.

🆔 @hackernewsgithubprojects

📦 drowzeys/keys-wasserman-slate-ai-studio-powered-by-tensorfold-mac-dgxspark

Keys Wasserman Slate AI Studio: Your Personal Film Crew

A single Mac can now function as a complete film production crew. Keys Wasserman Slate AI Studio merges a professional planning workflow with high-speed AI rendering, letting you script scenes, cast characters, and generate finished video clips with sound directly on your machine. The platform handles the entire pipeline by converting detailed text descriptions into visual assets using specialized models, effectively automating the transition from concept to final cut without needing a massive server cluster. It is a practical demonstration of how local hardware can manage complex creative tasks that once required cloud infrastructure, giving creators total control over their digital assets.

🆔 @hackernewsgithubprojects

📦 init-co-jp/diff-tour

How to Read Code Like a Senior Engineer

A command line tool now turns your git changes into a guided tour that mimics how a senior engineer actually reviews code. Instead of just staring at a wall of text, you step through a narrative that explains the purpose, points out risky spots, and asks you to predict outcomes before revealing the answer. It uses an AI model you already have to read your code, but it locks the AI down so it can only look and never touch your files. This is perfect for understanding complex refactors or reviewing AI-generated code where the logic is not obvious.

🆔 @hackernewsgithubprojects