📦 tencentarc/gae-geometricautoencoder

GAE: 3D Video Generation

GAE is the video generation model that finally bakes three-dimensional structure directly into every frame it creates. Most AI video tools just guess at shapes, which often leads to objects warping or disappearing as the camera moves, but this system takes a single photo and a text prompt to render a full video sequence paired with an actual three-dimensional point cloud. It achieves this by learning a hidden space that understands physical geometry, ensuring the virtual world looks consistent from every angle. It is a powerful step toward making synthetic video as reliable as real footage.

🆔 @hackernewsgithubprojects

📦 tencentarc/gae

GAE: The Geometry-Native Latent Space

Tencent ARC’s new project, GAE, turns a single image and a text prompt into a multi-view video that includes a perfectly consistent three-dimensional point cloud. Most generation tools create flat pixels, but this system learns a compact internal representation where geometry is built into the data from the start. It compresses frozen features into a state that decodes into both color and depth, ensuring the virtual world stays physically coherent no matter how the camera moves. This approach solves the problem of generated scenes that look visually rich but fall apart in three-dimensional space. By embedding structure directly into the latent space, it allows for stable, realistic world generation.

🆔 @hackernewsgithubprojects

📦 0xsero/sitegeist

Sitegeist: Your Browser's New Sidekick

Sitegeist is a browser extension that turns your computer into a collaborative workspace for an AI assistant. It lives in your sidebar, letting you guide it to click, type, and navigate any website to automate repetitive tasks or extract specific data. The coolest part is that it works in the background, so you can run multiple tasks at once without it stealing your focus or messing with your other tabs. Since it runs locally, your data stays on your machine, and you can bring your own API keys for providers like Claude or OpenAI. It’s not just a chatbot; it’s a tool that actually does the work.

🆔 @hackernewsgithubprojects

📦 clarkemedia/email-signature-generator

Email Signature Generator

An entire professional email signature builder fits inside a single file. No server, no account, and no tracking. Just a clean HTML page that runs entirely in your browser. It solves the annoying problem of getting stuck in complicated email settings by letting you pick fonts, add your logo, and tweak colors right before your eyes. The coolest part is the live preview. You see exactly how your signature looks in both light and dark modes instantly. If a field is empty, it simply vanishes from the output. It is a clever, private way to look professional without handing your data to a third-party app.

🆔 @hackernewsgithubprojects

📦 gitroomhq/postiz-agent

Postiz Agent

Your AI assistant can now cut your long YouTube videos into captioned, vertical clips and draft posts for them across twenty-eight platforms automatically. That is the headline feature of Postiz Agent, a command line tool designed specifically to let large language models like Claude handle your entire social media workflow. Instead of manually copying content between Twitter, LinkedIn, and TikTok, you simply ask your AI to schedule the post, upload the media, and handle the platform specific details for you. It turns your coding assistant into a capable social media manager. If you are tired of copy and paste, this is the automation you have been waiting for.

📰 https://news.ycombinator.com/item?id=49802697

🆔 @hackernewsgithubprojects

📦 vcruz305/qwen3.8-flash-next-exl3-dgx-spark-recipe

Run Qwen3.8 Flash Next on One Spark

Run Qwen3.8 Flash Next on a single NVIDIA DGX Spark without the usual server cluster headache. This repo is a practical guide for serving a massive, quantized large language model on that specific piece of hardware. It solves the tricky problem of making heavy AI models fit and run fast on a compact, unified memory system. You can serve it through a direct, highly tuned engine for maximum speed, or through a more standard API interface if you need broader compatibility. It is useful for developers who want real-world performance numbers and a clear setup path on this niche device.

🆔 @hackernewsgithubprojects

📦 hkustgz-ml4health-lab/apcyc

APCyc Designs Cyclic Peptides

APCyc automatically folds flat chains of amino acids into stable, ring-shaped drugs that can actually fit into specific biological targets. The tricky part of drug design is making sure these loops close up correctly while still binding tightly to their intended protein pockets. This tool uses a specialized AI model to predict exactly which atoms should connect and how the structure should twist, skipping the tedious guesswork of manual chemistry. By focusing on these precise molecular connections, it generates candidates that are structurally sound and biologically relevant. You get a smarter way to explore peptide designs that might otherwise be impossible to construct by hand.

🆔 @hackernewsgithubprojects

📦 copilotkit/openmuse

OpenMuse: The Personal Agent That Keeps Working

OpenMuse gives a computer to a personal AI agent that you can actually see working on your phone. The most striking part is not just that it browses the web, but that it can take over a live browser session if you need to intervene, then let it keep going. It’s essentially a digital intern that handles emails, runs terminal commands, and tracks goals while you watch it go. The project is built with CopilotKit and runs its own server, so you can dig into the source code and tweak the whole setup.

🆔 @hackernewsgithubprojects

📦 bevibing/tutor-skills

Tutor Skills: Your Personal AI Study Partner

Tutor Skills is the study assistant that turns your messy files into a personal quiz master. It takes any document, PDF, or codebase and builds a structured study vault inside your Obsidian app. The coolest part is that it doesn't just make notes; it actually quizzes you. It tracks exactly which concepts you know and which ones trip you up, then drills those weak spots until you get them right. It’s like having a patient tutor who knows exactly where you’re struggling. If you’re studying for an exam or learning a new codebase, this tool helps you focus your energy on what actually matters.

🆔 @hackernewsgithubprojects

📦 anthropics/claudeforfoundationmodels

Claude for Foundation Models: The Bridge You Didn't Know You Needed

Plug Claude directly into Apple’s new Foundation Models framework to drive it with the exact same session APIs you already use for on-device models. This package solves the headache of juggling separate clients for server-side logic by making Claude look like a native Apple model. The coolest part? If your chosen model politely refuses a tricky question, you can chain a fallback model that seamlessly picks up mid-sentence, so the user gets one smooth answer without even noticing the switch. It’s a clever way to keep your app logic simple while accessing Anthropic’s power. Stop writing custom glue code and start building apps that feel native to both worlds.

🆔 @hackernewsgithubprojects

📦 google-research/rrsi

RRSI: The Self-Improving Agent

Take a frozen AI model and watch it get smarter without changing a single weight. That is exactly what RRSI does. It treats the scaffolding around an agent as an evolving piece of software, tweaking prompts and memory until the system solves real coding and design tasks. The problem is simple: most agents forget how to handle new jobs after practicing on old ones. RRSI fixes this by adding strict rules that stop the system from cheating or overfitting. It is the closest thing to a self-improving robot we have, proving that better tools matter just as much as better brains.

🆔 @hackernewsgithubprojects

📦 nvlabs/grail

GRAIL

Train a robot to climb stairs without ever touching the real thing. That is exactly what NVIDIA’s GRAIL does. It takes 3D models and video clips to simulate humans moving and grabbing objects in the digital world. Then it teaches a robot brain to copy those moves. The coolest part? The robot can pick up items and climb actual stairs in reality, all from that fake data. It saves researchers from building a thousand test tracks just to learn one trick. If you want to know how robots learn to move without a physical teacher, this is the project to watch.

🆔 @hackernewsgithubprojects

📦 malevrigns/agent-jev

AgentJev: The AI Reflex That Doesn't Write Words

AgentJev is the tiny AI model that answers your questions without writing a single word. Usually, we ask big AI models to explain their decisions in full sentences, which is slow and messy. This project changes that entirely. You give it a messy state, like a stack trace or a code diff, and a simple question. Instead of typing out an answer, it runs through once and instantly gives you a probability for every option you listed. It’s like a reflex for your software, answering in milliseconds instead of seconds. It hits seventy-nine percent accuracy on the test set and saves a lot of time.

🆔 @hackernewsgithubprojects

📦 maneek21/deft

Deft: The Workspace Where AI Agents Work Like Colleagues

Deft lets you hire AI agents to handle real work right alongside your human team. Instead of treating AI as a separate chatbot, this self-hosted workspace puts agents into the same channels, tasks, and notes as your people. The coolest part is that agents like Defty can actually draft tasks for you based on your team’s shared knowledge. You review the proposal, approve it, and the task is created with full context attached. It’s a smart way to keep a clear record of what the AI did and why, giving you total control without losing the speed.

🆔 @hackernewsgithubprojects

📦 byk/loreai

Lore: The Memory That Compounds

Lore is the AI memory system that stops your coding agent from forgetting everything after the first hour. It sits quietly between your favorite coding tool and the model, acting like a smart secretary that remembers exactly which files you changed and why. Instead of just summarizing your chat into a vague summary that loses important details, it captures the specific facts and decisions you made. This means your AI can work on a project for days without getting confused or losing track. It even lets you review what your AI learned in a simple text file right inside your code folder.

🆔 @hackernewsgithubprojects

📦 ringlochid/oh-my-subagents

Oh My Subagents: The AI Team Manager

Oh My Subagents is the local orchestrator that finally stops you from babysitting AI agents. Instead of manually spawning and polling temporary tools, it manages them as a durable, accountable team with persistent state. The real magic is interruption recovery. If a session crashes or the internet blips, it doesn't start over from scratch. It remembers exactly where the work stopped and picks right back up. This turns fragile, one-off prompts into reliable, repeatable workflows. You get a structured team that actually finishes the job, not just a chat log that dies when the tab closes. It is the difference between playing with AI and actually managing it.

📰 https://news.ycombinator.com/item?id=49140789

🆔 @hackernewsgithubprojects

📦 unreallabsai/unreal-agent

unreal-agent

Build a coding agent that actually handles long-running tasks without freezing. Unreal Agent from Unreal Labs is an asynchronous framework that lets your AI run multiple commands simultaneously. Instead of waiting for one file read to finish before starting the next, it fires off independent jobs in parallel. This means your agent can inspect code, run tests, and process images all at once, dramatically speeding up complex workflows. It treats these tasks like background workers, so nothing blocks the main brain of your project. If you are tired of agents stalling on slow operations, this tool keeps everything moving. Grab the source code and start building smarter, faster software today.

📰 https://news.ycombinator.com/item?id=49805748

🆔 @hackernewsgithubprojects

📦 harnessrouter/systemoneharness

System One Harness: AI That Never Lies

System One Harness turns a new kind of AI model into a reliable worker that never invents facts. While most chatbots guess and hallucinate, this project forces the model to pick one specific action from a strict list of options, like choosing a button from a menu. It tracks its own confidence and refuses to act if it is unsure, meaning it will simply stop rather than make up an answer. This creates a perfect record of every decision, making it ideal for high-stakes tasks like banking or medical records where errors are not allowed. It is not about chatting; it is about precise, verifiable execution.

📰 https://news.ycombinator.com/item?id=49778358

🆔 @hackernewsgithubprojects

📦 awdqwdasdg/comfyui-spectrum-qwen2.1

comfyui-spectrum-qwen2.1

This ComfyUI plugin skips most of the heavy lifting in your AI image generator to cut render times by nearly three. Instead of calculating every single step from scratch, it learns the pattern of your current image and guesses the rest, letting you generate high-quality visuals in half the time without changing your settings. It works by predicting the hidden state of the diffusion model so you get faster results with fewer calculations, making it perfect for quick iterations. If you generate images with ComfyUI, this node is the fastest way to speed up your workflow while keeping quality high.

🆔 @hackernewsgithubprojects

📦 lkarlslund/laya.cpp

Run AI Decisions in Pure C++

Run complex AI decisions directly from C++ without Python overhead. Lay A C P P takes a modern language model and compiles it into a lightweight native executable that runs entirely on your GPU. It uses Vulkan and CUDA to process text with strict precision, ensuring every answer matches a Python baseline exactly. This is great because it strips away heavy dependencies, letting you serve fast, reliable logic directly from code. You get the speed of a native app with the accuracy of a standard model, all in one tiny tool.

🆔 @hackernewsgithubprojects