📦 michaelmoroz/slang-splat

Build Real-Time Gaussian Splatting with Slang Splat

Unlock high-performance 3D rendering with Slang Splat, a powerful Gaussian splatting trainer and viewer built on Slang compute shaders. This project provides a complete pipeline for rendering and training, featuring a real-time viewer, a COLMAP-based training system, and specialized support for custom scene initialization. It utilizes Vulkan for efficient GPU performance and includes advanced tools like photometric compensation to improve color accuracy across different camera views. Whether you need to load complex PLY scenes or run full training sessions, this repository offers a robust, highly modular framework for modern volumetric rendering. Dive in to start building your own splatting workflows.

🆔 @hackernewsgithubprojects

📦 zimengxiong/tabbed

Master Your Desktop Workflow with Tabbed

Are you tired of cluttering your screen with endless, overlapping windows while you work? Meet Tabbed, a clever macOS utility that brings a much-needed sense of order to your digital workspace. By simply holding the command key while dragging windows, you can effortlessly snap them together into organized tab groups, keeping your desktop clean and your focus sharp. It is built as a lightweight background application that integrates directly into your system, providing a seamless way to manage multiple windows at once. Give your productivity a boost and start tabbing your windows today for a truly stress-free computing experience.

🆔 @hackernewsgithubprojects

📦 rockxyapp/rockxy

Rockxy: The Open-Source macOS Proxy You Need

Stop struggling with network issues and start debugging like a pro with Rockxy, a powerful, open-source HTTP debugging proxy built natively for macOS. Designed by developers for developers, this tool lets you effortlessly intercept, inspect, and modify your HTTP, HTTPS, WebSocket, and GraphQL traffic. It solves the headache of hidden network calls by giving you deep visibility into your stack with features like request breakpoints, response mocking, and advanced filtering. Whether you are replaying API requests, throttling network conditions, or debugging with AI assistants, Rockxy provides a local-first, auditable workflow to help you build and test with total confidence.

🆔 @hackernewsgithubprojects

📦 dazuiba/handoff

Stop Overpaying for AI Coding Agents

Are you tired of expensive AI subscriptions burning through your budget? Meet Handoff, a tool that lets your coding agents collaborate seamlessly without constant manual switching. It solves the problem of high costs and fragmented context by delegating tasks to different models behind the scenes. You can send execution-heavy work like refactoring or testing to cost-effective models like DeepSeek, while keeping powerful models for complex decision-making. Everything runs in the background of your current session, and you can even track progress with interactive tools. Start delegating tasks today and make your AI workflow smarter, faster, and much cheaper.

🆔 @hackernewsgithubprojects

📦 tinyfatco/troublemaker

Turn Your AI Agent Into a Multi-Platform Powerhouse

Ever wanted your AI agent to live where you do, from Slack to Telegram and even your email? This project is a powerful, multi-platform runtime that liberates your AI agent, allowing it to manage files, execute bash commands, and handle tasks across various messaging services. It uses persistent memory files and autonomous skills to remember your preferences and reuse custom tools across different chat sessions. Whether you run it in a secure Docker sandbox or directly on your host, this runtime keeps your agent productive and connected. Take control of your automated workflows and let your agent work everywhere today.

🆔 @hackernewsgithubprojects

📦 yuan1z0825/nature-skills

Master Academic Writing with Nature Skills

Struggling to get your research published in top-tier journals? Meet the repository that is changing how scientists work. This project provides a collection of reusable AI agent skills specifically designed for high-impact academic workflows. Whether you need to conduct multi-source literature searches, polish your manuscript prose, create publication-grade scientific figures, or even draft patent applications, these modular tools help you maintain professional standards with ease. By standardizing complex tasks into structured, actionable steps, you can finally reclaim your time for deeper scientific discovery. Level up your research productivity and bring your work closer to publication today.

🆔 @hackernewsgithubprojects

📦 openmoss/awesome-wam

Master Embodied AI with World Action Models

Are you ready to explore the future of embodied AI? This repository is your essential gateway to World Action Models, which uniquely combine predictive world modeling with advanced action generation. It serves as a comprehensive, curated hub featuring the first systematic survey of this emerging field. You will find detailed breakdowns of various architectures, training data, and evaluation protocols designed to help you master these complex systems. Plus, the project includes structured blogs for each paper to simplify key concepts. Dive in today to stay ahead of the curve and transform how you understand autonomous robotic action planning.

🆔 @hackernewsgithubprojects

📦 xytom/coding-tools-mcp

Turn Your Local Editor into a Coding Powerhouse

Stop fighting with limited tools and start supercharging your local development workflow. This project is a specialized coding agent runtime that exposes essential local primitives directly to any standard protocol client. Instead of relying on clunky wrappers, you get direct control over file searching, structured code patching, git status tracking, and command execution, all while keeping your data securely within a defined workspace boundary. It is designed to be model-neutral and highly configurable, allowing you to run complex tasks with robust permission gates. Give your AI agent real, safe access to your local environment and watch your productivity soar.

📰 https://news.ycombinator.com/item?id=48594594

🆔 @hackernewsgithubprojects

📦 dagroup-pku/humannet

HumanNet: Scaling Embodied AI with 1 Million Hours of Video

Are you ready to see the future of robotics? Meet HumanNet, a groundbreaking research hub focused on scaling embodied AI through massive human-centric video learning. By curating a massive one-million-hour corpus of first and third-person footage, this project provides the essential data needed for robots to understand fine-grained human activities. It even includes StableVLA, a vision-language-action model for robust policy learning, and the extensive RoVid-X dataset for training physics-aware models. HumanNet demonstrates that egocentric video is a powerful, cost-effective substitute for limited robot data, paving the way for smarter, more capable robots. Explore this repository to revolutionize your next project.

🆔 @hackernewsgithubprojects

📦 agentic-in/inferoa

Take Control of Your AI Agents with Inferoa

Are your AI agents struggling to maintain focus during long tasks? Stop relying on simple prompting and start mastering loop engineering with Inferoa. This inference-native agent harness gives your AI a durable runtime, allowing it to inspect, edit, and verify complex work across recursive loops until the objective is proven. By optimizing your token usage through prefix-cache discipline and intelligent model routing, Inferoa keeps your context clean and your costs under control. Whether you are automating coding workflows or running research tasks, this powerful harness ensures your agents self-correct and deliver results effectively. Try it today to elevate your agentic workflow.

🆔 @hackernewsgithubprojects

📦 abhinavkale-dev/fynt

Build Powerful Automations with Fynt

Stop juggling manual tasks and start building custom automation pipelines with Fynt. This self-hostable platform provides a clean visual node editor where you can drag, connect, and configure complex workflows without the headache. Whether you need to chain together AI models, sync data between apps like Notion and Slack, or build conditional branching logic, Fynt makes it simple. It features live execution streaming through WebSockets and handles your credentials with secure encryption at rest. If you are looking for a flexible, open-source way to automate your daily operations, this powerful toolkit is the perfect place to start your journey.

🆔 @hackernewsgithubprojects

📦 smithhenryd/cgm

Calibrating Generative Models: Align Your AI Outputs

Ever wonder how to make generative models behave exactly how you want? This repository provides a powerful, general-purpose framework for fine-tuning generative models to meet specific distribution-level constraints. Whether you need to control the mean of black-box features with CGM or match full feature distributions using kCGM with MMD objectives, this tool offers a flexible base interface. By implementing simple sample and log probability methods, you can calibrate diverse models, from molecular design to protein structures. Dive into this codebase to gain precise control over your model outputs and bridge the gap between generation and your specific target statistics.

🆔 @hackernewsgithubprojects

📦 harryhsing/omniagent

OmniAgent: The Future of Efficient Video Perception

Tired of AI models wasting resources by watching every single frame of a long video? Meet OmniAgent, a breakthrough framework that completely changes how machines understand video. Instead of a passive, brute-force approach, this native omni-modal agent acts like a human, iteratively observing, thinking, and choosing exactly when to grab frames, audio, or clips. By focusing only on the information that matters, it slashes frame consumption by 73 percent while outperforming massive models. It is an end-to-end reasoning cycle that makes video understanding smarter and faster than ever before. Check out the project to see the new standard in efficient perception.

🆔 @hackernewsgithubprojects

📦 techenthusiast167/webrecon

Master Your OSINT Reconnaissance with WebRecon

Struggling to map out complex digital footprints for your security research? Meet WebRecon, a powerful open-source tool designed to automate web reconnaissance for cybersecurity professionals. Unlike simple crawlers, this tool specializes in intelligence gathering, from harvesting emails and identifying technology stacks to discovering cloud storage. Its standout feature is the advanced relationship graph visualization, which maps connections between domains, profiles, and infrastructure into an interactive network map. Whether you are performing attack surface management or deep-dive investigations, WebRecon turns raw data into clear, actionable visual insights. Start mapping your targets smarter today and take your security analysis to the next level.

🆔 @hackernewsgithubprojects

📦 hp-luo/stare

Stop Model Entropy Collapse with STARE

Are you tired of your reasoning models losing their edge during training? Reinforcement Learning with Verifiable Rewards often suffers from policy entropy collapse, where models stop exploring effectively. STARE, or Surprisal-guided Token-level Advantage Reweighting, solves this by identifying entropy-critical tokens and selectively reweighting their advantages to sustain exploration. It uses a smart closed-loop gating mechanism that intervenes only when necessary to keep training stable. By effectively managing entropy during complex reasoning tasks, STARE consistently outperforms standard baselines like GRPO. Whether you are working with small or large models, STARE helps maintain the vital balance between exploration and exploitation.

🆔 @hackernewsgithubprojects

📦 mingyuchoi-run/lsm

Boost Image Quality with Linear Recurrent Units

Ever wonder how to sharpen blurry images while keeping computation efficient? This repository features an innovative approach to image super-resolution using a Linear Recurrent Unit combined with a special semantic modulating unit. By integrating spatial categorization and feature enhancement, this method achieves state-of-the-art results that balance performance with efficiency. Whether you are dealing with classical or lightweight super-resolution tasks, this project offers the official code and pretrained models needed to elevate your visual processing projects. Dive into the code to see how these advanced techniques can help you achieve clearer, more detailed images today.

🆔 @hackernewsgithubprojects

📦 gy65896/cotir

Fix Any Image with CoTIR: Internalized Chain-of-Thought Restoration

Struggling to fix complex, degraded photos? Most all-in-one restoration models fall short when things get complicated, but there is a better way. Meet CoTIR, a breakthrough framework that brings Internalized Chain-of-Thought reasoning to image restoration. By treating restoration as a specialized editing subtask, it optimizes results using pre-trained models without the need for slow, multi-step processing. It effectively encodes structured reasoning directly into its learning objective, leading to superior perceptual quality across various scenes. Whether you are dealing with mixed noise or blur, CoTIR provides a powerful, holistic solution for cleaner, sharper images. Experience the next level of image recovery today.

🆔 @hackernewsgithubprojects

📦 malik-group/do-as-i-do

Teach Robots Human Skills from Video!

Ever wondered how to teach a robot to mimic human actions? This repository makes it possible by converting simple hand-object demo videos into precise robot motion trajectories. It breaks the complex process into a multi-stage pipeline that starts by reconstructing 3D hand and object meshes from your footage. It then retargets those movements onto a robot hand using physics-based optimization to ensure the motion is physically consistent. Finally, you can stream these trajectories directly to real hardware. Whether you are into robotics or computer vision, this project offers a fascinating way to turn human demonstrations into actionable robotic tasks.

🆔 @hackernewsgithubprojects

📦 mingyuee88/sae-post-intervention-recovery

Are LLM Interventions Actually Reliable?

Ever wonder if suppressing model behavior through latent interventions is as secure as we think? This project dives into the reliability of Sparse Autoencoder feature clamps by testing whether suppressed behaviors can be recovered from the residual state without reversing the initial intervention. Using tools like encoder-orthogonal projection and cross-layer Jacobian methods, the researchers investigate if these interventions truly form effective behavioral bottlenecks. It covers various settings including refusal, unlearning, and circuit analysis. Explore the core utilities provided in the repository to evaluate your own safety interventions and see if your model defenses are truly holding firm.

🆔 @hackernewsgithubprojects

📦 comfy-org/comfy-kitchen

Supercharge Diffusion Models with Comfy Kitchen

Are your diffusion models running slower than you want? Meet the library that acts as a high-performance kernel engine for your inference tasks. This project provides optimized kernels specifically for diffusion workflows, supporting multiple backends like CUDA and Triton to maximize your hardware potential. By using specialized tensor subclasses, it automatically selects the fastest available backend for your operations, handling complex quantization tasks effortlessly behind the scenes. Whether you are running on Blackwell or older architectures, it optimizes your compute paths for peak efficiency. Give your AI projects a significant speed boost and optimize your inference pipelines today.

🆔 @hackernewsgithubprojects