📦 kxding/playworld

PlayWorld: The AI Agent That Actually Plays Games

PlayWorld is a benchmark that finally fixes how we test AI video generators by letting an agent figure out its own moves instead of forcing it to follow a script. Instead of checking if a model can replicate a predetermined path, this project uses a smart controller that watches the generated video, understands the goal, and decides what to do next. It tests whether these video models really understand physics and consistency over time by seeing if the agent can successfully complete complex tasks. This reveals that most current systems still struggle with long-term logic, making PlayWorld the ultimate stress test for whether AI can actually play, not just predict.

🆔 @hackernewsgithubprojects

📦 amap-ml/dreamx-phi

DreamX-Phi Video World Model

DreamX-Phi turns a simple picture and a list of robot movements into a realistic video of the future. It is a video world model designed for robotic manipulation, meaning it predicts exactly how an object will move and look after you tell a robot how to grab it. By taking an initial photo and specific action steps, the system generates a smooth clip that respects real physics, keeping the scene stable and the object moving correctly. This helps researchers test robotic strategies without wasting time or breaking expensive hardware. It is a clever way to simulate robotic actions visually before the real robot ever lifts a finger.

🆔 @hackernewsgithubprojects

📦 aerovato/nitro

Nitro: Type commands in plain English

Nitro turns your terminal into a conversational assistant that translates everyday English into actual shell commands. You type something like find all markdown files and count lines and it builds the complex bash command for you instead of guessing syntax. It connects to your favorite AI providers like OpenAI or Anthropic to handle the thinking. There is a built safety system that grades commands by risk and asks for approval before it deletes or overwrites anything. You can run it once or stay in a chat mode to keep asking follow-up questions. It is a simple way to save time on repetitive terminal tasks without memorizing obscure flags.

🆔 @hackernewsgithubprojects

📦 garystack/parallel-rl

Parallel RL Merging

Parallel RL solves the messy problem of training AI models on multiple skills. Usually, teaching a model math then code causes it to forget the math, a phenomenon called catastrophic interference. This project reveals that reinforcement learning updates are actually sparse and point in different directions, allowing you to train each skill separately and then simply add them together. The repository provides the code to analyze these update patterns and merge the resulting model adapters without retraining from scratch. You get a single, capable model that handles multiple tasks cleanly. It is a clever way to combine specialized skills into one generalist.

🆔 @hackernewsgithubprojects

📦 ibil-code/chis

CHIS

Synthesize realistic medical tissue slides without spending a single second training or fine-tuning a massive model. The chis project does exactly that by taking a simple layout mask of how you want the cells arranged, blending it with the textures of a reference image, and generating highly realistic histopathology images. It clever-hacks pretrained diffusion models using frequency and wavelet tricks to keep the structure intact while borrowing the perfect visual style. Best of all, it is lightweight enough to run on a consumer graphics card right at home. It is a brilliant shortcut for generating diverse, high-fidelity medical images on demand.

🆔 @hackernewsgithubprojects

📦 lmresiliency/lm-resiliency

lm-resiliency

lm-resiliency is the training safeguard that saves your massive language model from crashing without wasting days of compute. When you train huge models across many GPUs, hardware glitches and memory errors are inevitable and usually mean you have to restart your entire job from scratch. This tool solves that by saving complete snapshots of your training state directly into memory every few steps instead of just writing them to slow disk. If a GPU fails or a network cable gets pulled, the system instantly grabs the latest saved state from a neighbor machine and jumps back in, losing only minutes of work instead of weeks.

🆔 @hackernewsgithubprojects

📦 dagroup-pku/physisforcing

physisforcing

The training tool that makes robotic video generation actually obey the laws of physics. Physisforcing is an add-on framework that helps existing AI video models generate physically realistic simulations for robot training. Instead of letting the AI hallucinate random movements, it focuses on the exact points where objects touch and interact, aligning both pixel movement and the overall relationships between objects. This simple tweak makes the simulated environments much more realistic, helping virtual robots learn tasks and successfully plan actions in the real world. Check out the project page to see these physics-accurate robotic simulations in action.

🆔 @hackernewsgithubprojects

📦 glitetech/glite-arf

glite-arf

Autonomous AI agents are incredible at writing code and analyzing data, but when left to run research projects on their own, they quickly create a chaotic mess of unread prose and broken pipelines. The glite-arf repository solves this by wrapping AI research in a strict, file-based framework of hard structural guardrails. Every single AI task is completely isolated in its own folder, and automated scripts verify every file structure and metric before anything can be committed. It even locks past completed work so later AI runs can build on top of frozen results without breaking them, keeping your automated research clean, organized, and reproducible.

🆔 @hackernewsgithubprojects

📦 z-yiwei/rectify-then-diffuse

Rectify Then Diffuse: Fixing AI Image Concepts

Rectify Then Diffuse treats complex image generation like a simple math problem by fixing the starting point before the noise removal begins. The project solves the common issue where AI tools blend two distinct ideas into one confused mess. It works by running a quick preview to spot overlapping features, then gently adjusts the initial image data so each concept claims its own space. This means you can ask for a red backpack and orange glasses, and get exactly that without the colors bleeding together. The approach skips heavy training, making it a clever, lightweight trick that keeps separate objects clear and distinct.

🆔 @hackernewsgithubprojects

📦 yaxin9luo/autodesign

AutoDesign: One Paper, Four Outputs

AutoDesign turns a single academic PDF into four distinct, editable artifacts. It generates a dense poster, a full presentation deck, an editorial web page, and even a narrated video. Instead of just writing code, it optimizes a reusable design system that learns from its own mistakes. It starts with your paper, then iteratively refines the layout, typography, and content until the results look professional. The best part is that everything stays editable, so you can tweak the slides or poster without starting from scratch. It essentially automates the entire afterlife of a research paper. Check it out if you need to make your work look as good as it reads.

🆔 @hackernewsgithubprojects

📦 nousresearch/hermes-bot-mode

Hermes Bot Mode: Create Your Own AI Agent Team

Hermes Bot Mode lets you run multiple distinct artificial intelligence personalities simultaneously, each with its own chat history and schedule. This desktop plugin transforms your agent profiles into a roster of named bots, complete with unique avatars and daily routines. The most surprising feature is bot-to-bot messaging, where agents can send messages to one another using a simple inbox system. You can type a mention in any chat to have your active bot hand off work to a specialist, wait for the reply, and report back. It works like a team of digital assistants collaborating in the background without needing extra servers or complex setups.

🆔 @hackernewsgithubprojects

📦 garystack/beyond-mmenv-scaling

Beyond-MMEnv-Scaling: Smarter Training Data for AI Vision

Beyond-MMEnv-Scaling is the research project that finally teaches AI vision models how to choose their training diet rather than just eating more of it. Most developers assume that throwing more visual environments at a machine learning model will automatically make it smarter, but this code proves that is usually a waste of time. Instead it uses two clever tricks. First it picks environments based on the actual skills they practice so the agent learns distinct abilities instead of repeating the same pattern. Second it arranges those tasks into a structured curriculum that slowly ramps up the difficulty.

🆔 @hackernewsgithubprojects

📦 harshkedia177/axon

Axon: The code knowledge graph

Axon is the code intelligence tool that turns your entire codebase into a searchable knowledge graph for both developers and AI agents. Instead of just reading flat text, it builds a visual map of every dependency, call chain, and execution flow in seconds. You can explore this map interactively in a web dashboard or let your AI assistant query it directly to understand exactly what will break before you make a change. It runs completely locally without any cloud dependencies, giving you a clear, structural view of your architecture. This approach saves you from blind edits and helps you spot hidden connections that standard search tools completely miss.

🆔 @hackernewsgithubprojects

📦 zzaiyan/geoseg-ov

GeoSeg-OV: Smarter Satellite Segmentation

GeoSeg-OV is the remote sensing tool that finally helps computers understand satellite maps across different places without retraining from scratch. Most AI models for analyzing aerial images get confused when the scenery changes, but this project uses structural hints from depth models to guide the recognition process instead of just matching pixels. It keeps the text understanding intact while letting the computer see where objects actually sit in space. This makes it surprisingly good at spotting land types like forests or roads in completely new regions you haven’t seen before. It is a clever way to bridge the gap between different geographic datasets.

🆔 @hackernewsgithubprojects

📦 zebinx/drivevla-m0

DriveVLA-M0

Autonomous cars can now remember their past mistakes to drive safer. DriveVLA-M0 is a research project that builds a memory for self-driving systems. Instead of just reacting to traffic right now, it stores detailed records of near-crashes and confusing road situations. When the car encounters a similar problem later, it quickly pulls up that old memory to correct its path. This failure-aware system helps vehicles learn from history, reducing accidents by anticipating trouble before it happens. It is a smart step forward for safer, smarter autonomous driving.

🆔 @hackernewsgithubprojects

📦 luo-z13/eo-wm

eo-wm

Satellite observations can now predict how vegetation degrades under extreme weather by treating the planet like a physics-driven video game. A new project called eo-wm uses a latent video diffusion model to forecast earth observations from sparse satellite data and weather patterns. By splitting weather data into baseline climatology, anomalies, and cumulative stress signals, the model creates realistic, probabilistic forecasts of environmental changes. It even includes benchmarks to test how plants respond to severe summer heat and drought. Ultimately, this approach turns raw climate data into highly responsive, visual simulations of our changing planet.

🆔 @hackernewsgithubprojects

📦 dosu-ai/decant

Decant: Your AI Session Dashboard

Decant is the local privacy tool that turns your Claude Code and Codex session logs into a searchable dashboard showing exactly where your tokens, cost, and agent time go. It runs entirely on your machine without sending any data out, letting you track spending, inspect context windows, and find which files your AI assistants actually touched. It gives you clear, tangible insights into your coding agent habits without the privacy risk of cloud-based analytics.

📰 https://news.ycombinator.com/item?id=49279140

🆔 @hackernewsgithubprojects

📦 jakubantalik/libraries

Libraries

Build beautiful interfaces without writing complex CSS animations yourself, because this library handles the heavy lifting for you. Think of it as a magic wrapper for your React components that instantly adds stunning visual effects. You get a traveling beam of light that glides around your buttons and cards, or a breathing halo that pulses softly behind your content. The best part is how smart it is, it automatically detects the shape of your elements and adjusts the glow perfectly. It even respects your system settings, turning off fancy animations for anyone who prefers reduced motion, which is incredibly thoughtful for accessibility.

🆔 @hackernewsgithubprojects

📦 deepseek-ai/awesome-deepseek-agent

Awesome DeepSeek Agent: The Ultimate Integration Guide

Awesome DeepSeek Agent is a curated collection of step-by-step guides that lets you plug DeepSeek models into your favorite coding and agent tools. Instead of hunting through scattered documentation, you get clear instructions for setting up DeepSeek-V4-Pro or Flash in twenty different applications like Cline, GitHub Copilot, and Cherry Studio. The project solves the headache of configuration by covering installation, API key setup, and initial runs in plain language. It is particularly useful because it highlights powerful features like a one million token context window and adjustable reasoning effort levels that many users miss.

🆔 @hackernewsgithubprojects

📦 jlcodes99/cockpit-tools

Cockpit Tools: The Multi-AI IDE Account Manager

Cockpit Tools is the universal account manager that finally unifies all your AI coding assistants into one single dashboard. Instead of juggling logins across Cursor, GitHub Copilot, Windsurf, and half a dozen other platforms, this tool lets you group every account into neat categories and switch between them with one click. The real magic happens with quota monitoring, where you can instantly see how many API calls you have left across all your services without visiting each website manually.

🆔 @hackernewsgithubprojects