📦 dang-ai/llm-training-holistic-data-schedule

Optimizing LLM Pre-training with Holistic Data Scheduling

Holistic Data Schedule is the advanced framework that applies multi-objective reinforcement learning to optimize data selection for language model pre-training. By dynamically scheduling data, it addresses the complexities of training large models, specifically improving efficiency and downstream task performance on benchmarks like MMLU. The repository provides the complete implementation, including specialized CUDA kernel configurations, data preprocessing scripts for the Pile dataset, and comprehensive training routines. Researchers can leverage these tools to fine-tune data mixtures during the training process, ultimately leading to more robust model development and clearer insights into how different data domains impact overall accuracy.

🆔 @hackernewsgithubprojects

📦 frontisai/naturebench

Can AI Agents Beat Scientific Breakthroughs?

A terminal window executes a command that automatically downloads scientific datasets and launches containerized environments to test AI coding performance. NatureBench provides a standardized cross-discipline benchmark featuring ninety tasks derived from peer-reviewed scientific papers. It evaluates whether AI agents can actually solve complex machine learning problems rather than just reproducing existing code. By comparing agent results against reported state of the art benchmarks, researchers can identify if AI models are capable of moving toward true scientific discovery. This project offers a rigorous framework for testing and scoring agent performance across diverse scientific domains.

🆔 @hackernewsgithubprojects

📦 little-book-of/c

The Little Book of C: A Masterclass in Systems Programming

The Little Book of C is the technical guide that finally demystifies the foundational language of modern computing through a clear, structured format. This resource provides a comprehensive walkthrough of C, moving through sequential sections that explain core concepts while offering practical source code examples for every lesson. By focusing on the essential building blocks, it helps developers master the language that powers operating systems and hardware worldwide. Whether you are learning to manage memory or refining your systems programming skills, this repository serves as a portable, reliable handbook for becoming a more effective programmer.

🆔 @hackernewsgithubprojects

📦 refloow/refloow-geo-forensics

Analyze Image Locations Privately with Refloow Geo Forensics

Analyze and visualize geographical data from your images and videos by using this open-source digital forensics tool. It automatically extracts EXIF metadata from batch files and plots GPS coordinates directly onto interactive maps to help you track movements or verify events. Because it operates entirely locally on your machine, it ensures your sensitive investigative data remains private without requiring any cloud uploads or logins. This lightweight application reconstructs chronological timelines from media files, making it a powerful resource for security analysts and researchers needing a fast, efficient way to uncover hidden connections in visual evidence.

🆔 @hackernewsgithubprojects

📦 gastruc/universat

Process Any Earth Observation Data with One Model

Analyze diverse Earth observation data using a single transformer model capable of handling varied resolutions, sensors, and modalities without the need for manual resampling or channel selection. By employing a Universal Patch Encoder, this project maps different spatial, spectral, and temporal inputs into a shared embedding space, allowing the same model to process everything from high-resolution optical imagery to radar time series. It utilizes sub-patch skip connections to preserve fine spatial details at any requested output grid. This approach enables consistent, high-quality feature extraction across global datasets, offering a versatile foundation for advanced remote sensing tasks and monitoring applications.

🆔 @hackernewsgithubprojects

📦 lancezpf/agent-as-a-router

Optimize Coding Tasks with Agent-as-a-Router

Optimize your coding workflows by routing tasks to the best backend models based on performance and cost. This repository provides a complete implementation for managing model decisions, helping you balance quality with expense effectively. It includes the CodeRouterBench dataset for evaluation, along with tools to reproduce research results offline. You can easily integrate the router into your own systems using the provided Python pipeline and configuration scripts to handle complex coding problems through a verifier-backed feedback loop. Leverage these official artifacts to experiment with model routing and refine your development pipeline for better efficiency.

🆔 @hackernewsgithubprojects

📦 joeynyc/spark-doctor

Debug your NVIDIA DGX Spark with Spark Doctor

Running a diagnostic scan shows a clean terminal output detailing system health, GPU metrics, and Docker configurations. This tool serves as a local diagnostic CLI for the NVIDIA DGX Spark platform, gathering scattered performance data into one command to pinpoint issues like power caps, thermal risks, or memory pressure. It validates your model recipes against hardware constraints and provides actionable, plain-English instructions for troubleshooting rather than automated fixes. This ensures you understand exactly why a failure occurred and how to resolve it safely without hidden changes to your setup. Use it to keep your hardware running reliably.

🆔 @hackernewsgithubprojects

📦 inclusionai/sing-guard

Build Dynamic AI Safety with SingGuard

A developer runs a simple Python script and receives a clear safe or unsafe judgment alongside a specific risk category for a harmful user query. SingGuard provides a flexible, policy-adaptive multimodal guardrail model that evaluates text, images, and cross-modal content for safety. Instead of relying on static training rules, it allows you to inject custom natural-language policies at runtime, ensuring your safety checks stay relevant to specific needs. By supporting fast-to-slow dynamic reasoning, this tool helps you efficiently moderate content in production environments, making it a powerful utility for keeping AI interactions secure and aligned with your unique standards.

🆔 @hackernewsgithubprojects

📦 cisco-ai-defense/skill-scanner

Automated Security Scanning for AI Agent Skills

Skill Scanner provides automated security analysis for AI agent configurations by detecting threats like prompt injection, data exfiltration, and malicious code patterns. This tool employs a multi-engine approach, combining pattern-based detection with behavioral dataflow analysis and LLM-as-a-judge capabilities to identify vulnerabilities while filtering out noise. It is designed for integration into development workflows, supporting CI/CD pipelines through reusable actions and pre-commit hooks. While this software offers extensive scanning coverage to help mitigate risks, it serves as a best-effort defense layer that should be paired with human review to ensure robust security in production environments.

📰 https://news.ycombinator.com/item?id=48656076

🆔 @hackernewsgithubprojects

📦 halilkirazkaya/arsenal-ng

Modern Pentest Command Launcher for Linux and macOS

This command launcher provides instant access to over two hundred cybersecurity cheat sheets directly within your terminal. Built in Go, it features a fast, responsive interface that allows for smart fuzzy searching across command names, descriptions, and tags. The tool simplifies complex workflows by supporting global variables that automatically populate command arguments, reducing repetitive typing during engagements. Once a selection is made, the command is written directly to your terminal buffer, ready for immediate execution. It acts as a lightweight, efficient reference library that keeps your security testing commands organized, searchable, and always at your fingertips.

🆔 @hackernewsgithubprojects

📦 ccbkkb/microwarp

Run a Cloudflare WARP Proxy with Only 800KB of RAM

A terminal window displays a Docker stats report showing the container running with merely 800 kilobytes of memory usage while routing network traffic. MicroWARP is an ultra-lightweight SOCKS5 proxy built for servers that replaces memory-heavy userspace clients with a native Linux kernel interface and a pure C engine. By utilizing WireGuard at the kernel level, it achieves near-zero CPU overhead and functions seamlessly in resource-constrained environments. This tool is perfect for routing API traffic or obfuscating your server IP without consuming system resources. Try this minimalist solution if you need an efficient, high-performance egress gateway for your containers.

🆔 @hackernewsgithubprojects

📦 youmind-openlab/nano-banana-pro-prompts-recommend-skill

10,000 AI Image Prompts Right Inside Your Code Editor

This AI agent skill brings a massive library of over 10,000 curated image generation prompts directly into tools like Claude Code and OpenClaw. Instead of searching endlessly for the right style, you simply describe your needs or paste your content, and the skill automatically retrieves top-tier prompt matches along with sample images to guide your creative process. It features a smart remix mode that personalizes templates to fit your specific scripts or articles perfectly. Because the repository syncs with community-driven data twice daily, you always have access to current, high-quality styles for your projects.

🆔 @hackernewsgithubprojects

📦 bozhoudev/codex-orange-book

Mastering Codex with the Orange Book

A user scrolls through a detailed digital document that breaks down the entire process of installing and using the Codex platform from start to finish. This repository serves as a comprehensive guide that functions as an unofficial manual for those looking to master the technology through practical real world examples. By offering clear pathways from initial setup to hands on implementation, it demystifies complex workflows and provides a downloadable resource for developers. It is a valuable tool for anyone needing a structured reference to improve their technical proficiency. Grab this guide to streamline your learning and build projects more effectively.

🆔 @hackernewsgithubprojects

📦 jamesstringer90/appsandbox

Run Fully Sandboxed Virtual Machines on Windows and macOS

Build and run full desktop virtual machines like Windows, Ubuntu, and macOS using this high-performance, open-source tool. It offers an easy way to isolate your work by running programs or testing software inside clean, GPU-accelerated environments that remain completely separate from your host system. You can manage everything through a simple graphical interface or automate the entire VM lifecycle using a scriptable headless API and Python SDK. Whether you are testing drivers, running computer-use AI agents, or just need a secure sandbox for development, this project provides a flexible and powerful way to handle your virtualized tasks.

🆔 @hackernewsgithubprojects

📦 davideast/inbrowser-agent

Build AI Directly in the Browser

Developers can now run language models directly in the user's browser without needing a dedicated inference backend by utilizing the in-browser AI stack. This modular framework consists of four libraries that handle on-device model execution on the GPU, manage durable streaming jobs that survive page reloads, and provide a runtime for agents that can execute complex tool-based tasks. Whether you are building local AI tools or need to proxy cloud-based models while keeping API keys secure on the client, this project enables robust, persistent AI experiences that remain functional even when network connections drop or browser sessions are interrupted.

🆔 @hackernewsgithubprojects

📦 dmtrkovalenko/ffs

Search Your Disk Directly By Bypassing the Kernel

This command line tool performs file searches by bypassing the operating system kernel entirely and reading raw disk blocks directly. By avoiding the standard virtual file system and buffered read paths, it functions as a specialized file searcher that can scan detached volumes like disk images without mounting them. While it requires elevated permissions to access raw device nodes, this approach can outperform conventional tools on extremely large file sets by eliminating kernel overhead. It supports specific file system formats and utilizes multi-core processing to scan through massive data volumes, demonstrating that direct hardware interaction can sometimes offer significant performance advantages.

📰 https://news.ycombinator.com/item?id=48622433

🆔 @hackernewsgithubprojects

📦 juliensimon/canopy

Canopy: Parallel Claude Code Sessions for macOS

Canopy is a native macOS application that enables you to run multiple Claude Code sessions in parallel within a single interface. By leveraging git worktrees, it allows you to maintain independent coding contexts for different branches without the friction of stashing changes or juggling multiple terminal windows. The application includes a unified activity dashboard for monitoring token usage, a command palette for rapid navigation, and automated session management that persists your conversation state across restarts. Built with SwiftUI, it provides a high-performance alternative to managing disjointed AI workflows, offering a streamlined path to ship your projects faster while maintaining full control.

📰 https://news.ycombinator.com/item?id=48538315

🆔 @hackernewsgithubprojects

📦 kevinywu/hug

Generate Human Grasps from Stereo Images

Generate diverse human hand grasps for objects by simply clicking on an image captured from a stereo camera. This project provides an official implementation for Human Universal Grasping, a system trained on real-world data to predict precise grasp patterns in MANO form. It includes an interactive application that lets you visualize 3D hand poses and skeletons directly in your web browser. By processing RGB-D inputs, the model understands spatial geometry to create realistic grasping demonstrations. Explore this implementation to advance your work in hand-object interaction and robotics with a ready-to-use inference pipeline for your own captures.

🆔 @hackernewsgithubprojects

📦 stepfun-ai/gelab-zero

Build Your Own Local GUI Agent with GELab-Zero

Automate mobile tasks locally by deploying this open-source GUI agent solution that removes the need for cloud dependencies and complex infrastructure. Designed for both personal and enterprise use, it features a lightweight 4B model that runs on consumer hardware to maintain complete privacy. The repository includes a plug-and-play engineering pipeline to handle device management, task distribution, and interaction trajectories across various Android apps. By automating real-world tasks like shopping or navigation, this tool helps developers quickly test and deploy agent workflows. Download the project to regain control over your mobile automation tasks and experience fully local inference.

📰 https://news.ycombinator.com/item?id=48657446

🆔 @hackernewsgithubprojects

📦 xiangchenyin/holo-world

Holo-World: Controllable Video Generation from a Single Image

Holo-World is a unified video world model that enables precise camera, object, and weather control starting from just a single input image. Traditional models often struggle to manage these elements independently or rely on pre-existing source video structure, but this project introduces a specialized dataset called HoloStateData and a Unified Scene Adapter to resolve these limitations. By factorizing world preservation and environmental changes into distinct parameter subspaces, the model maintains scene structure while rendering specific weather effects and particle motion. This approach provides a robust framework for controllable video generation, offering a sophisticated tool for modern visual AI research.

🆔 @hackernewsgithubprojects