📦 tsinghua-fib-lab/omniscientist

OmniScientist

Artificial intelligence is transitioning from a basic writing assistant into a fully automated researcher capable of proposing its own scientific hypotheses. This is the core vision behind omniscientist, an open research ecosystem developed by Tsinghua University and the Zhongguancun Academy. The project connects large language models to scientific concept networks, allowing them to spot hidden connections between theories and propose genuinely novel research ideas. Beyond brainstorming, it automates experiment design by searching for codebases and datasets to build real testing plans. It is a fascinating glimpse into how human and machine researchers might soon collaborate to solve complex global challenges.

🆔 @hackernewsgithubprojects

📦 openbmb/minicpm-desk-pet

MiniCPM Desk Pet

MiniCPM-Desk-Pet is the desktop companion that brings a local AI assistant right to your screen as an animated pet. After a simple guided setup, the app downloads a lightweight language model that runs entirely on your own computer, so your conversations stay private and offline. Instead of just sitting there, this little pet actively monitors your coding workspace. It automatically notices when tools like Cursor or Claude Code are busy, displays custom animations to match their progress, and even rings a bell when an agent is waiting for your feedback. It is a brilliant way to make local AI feel alive.

📰 https://news.ycombinator.com/item?id=48824849

🆔 @hackernewsgithubprojects

📦 microsoft/healthagentbench

HealthAgentBench

Drop your AI agents into a terminal environment and see if they can actually survive real-world healthcare challenges. With healthagentbench, you can put frontier AI models to the test across fifty-four intense clinical and biomedical scenarios. We are talking about critical tasks like scanning chest X-rays, flagging errors in messy electronic health records, and matching patients to life-saving clinical trials. This is not a simple multiple-choice exam; the agents must inspect raw medical data, use actual terminal tools, and solve complex problems under realistic constraints. It turns out medicine is incredibly tough, as even the smartest models only solve a minority of these tasks. Grab the code and see how...

🆔 @hackernewsgithubprojects

📦 weiyixue999/signal_structure_aware_gaussian

Signal Structure-Aware Gaussian Splatting

Reconstructing massive 3D environments from photos usually crushes consumer hardware under the weight of massive image files and runaway memory usage. A new framework called signal_structure_aware_gaussian changes that by dynamically adjusting image resolution during training to match how stable the 3D scene is at any moment. It starts with lower-detail images to build a coarse layout, partitions the scene into blocks, and only dials up the resolution when the local geometry stabilizes. This smart scheduling prevents the system from creating redundant 3D points too early. The result is a highly efficient pipeline that lets developers reconstruct large-scale scenes on standard hardware in just over an hour.

🆔 @hackernewsgithubprojects

📦 xcltql666/omintryon

OmniTryOn

OmniTryOn is the video-generating tool that finally lets you swap outfits on moving people in a single step. Instead of spending hours meticulously editing frames or trying to warp flat images of clothes onto a moving body, this clever tool lets you virtually dress anyone in a video with absolutely any garment. It works by treating your clothing items as prompts, injecting them directly into the video generation process to keep everything looking natural, even when the person moves around. It is an incredibly smart way to see how clothes actually move and fit on a real, walking body before making a purchase.

🆔 @hackernewsgithubprojects

📦 sunrainyg/randopt

RandOpt

Pretrained language models are already surrounded by a dense thicket of highly specialized task experts just waiting to be uncovered through random noise. This is the core discovery behind randopt, an open-source codebase that introduces a surprisingly simple way to post-train AI models. Instead of relying on complex, mathematically heavy optimization steps, this approach samples random perturbations around pretrained weights to find highly diverse, top-performing task experts. It turns out that a single random step actually has a high chance of outperforming a traditional gradient update on strong models. Once these diverse experts are discovered, they can be distilled back into a single powerful model. It is a fascinating rethink...

🆔 @hackernewsgithubprojects

📦 26worldcup/26worldcup.github.io

26worldcup.github.io

The 26worldcup.github.io companion is the ultimate fan-built web app that keeps you ahead of every kick-off without any of the typical internet clutter. It serves up a beautifully clean schedule, group standings, real-time weather forecasts, and broadcasting guides for thirty-two countries. But the best part is the built-in tournament simulator. Driven by an Elo mathematical model built on forty-nine thousand historic international matches, you can instantly simulate the entire tournament up to ten thousand times to see exactly how your favorite team is predicted to finish. It runs fully offline, contains zero ads, and is translated into twenty-three languages. Check out this incredibly smart tool before the tournament begins.

🆔 @hackernewsgithubprojects

📦 fluidifyai/regen

Regen

Regen is the open-source on-call manager that finally lets you run your own alert and incident response system without the massive subscription bills. It serves as a self-hosted alternative to PagerDuty, pulling in alerts from tools like Prometheus, Grafana, and AWS to coordinate your team's response. The best part is its built-in artificial intelligence helper. You can hook up your own AI keys to automatically draft post-mortems, write incident handoffs, and summarize chaotic Slack threads. It is incredibly fast, manages schedules, and fits perfectly into your existing chat channels to keep everyone in the loop during a crisis mode.

🆔 @hackernewsgithubprojects

📦 boredos/boredos

BoredOS

Build and run your own software on a custom-made operating system written entirely from scratch. This project is a modern, hobbyist Unix-like operating system that features a layered window manager, multi-core processing, and its own graphics compositor. It goes far beyond standard terminal-based hobby systems by delivering a full user-space graphical desktop environment alongside a robust TCP/IP networking stack with built-in command-line tools like curl and ping. It is a masterclass in low-level engineering, showing you exactly what it takes to boot physical hardware, manage system memory, and draw pixels to a screen with absolute control.

🆔 @hackernewsgithubprojects

📦 dannymac180/fable-advisor

Fable Advisor

Running your coding sessions entirely on top-tier artificial intelligence models is wasting your budget on basic typing tasks. The fable-advisor plugin for Claude Code solves this by turning your development environment into an efficient, cost-saving assembly line. It positions Anthropic's most capable model as a high-level architect that handles the planning, specifications, and quality checks, while delegating the actual coding to much cheaper models. This smart routing of tasks cuts your token costs by roughly sixty percent while actually delivering higher quality code. It even includes a dedicated cross-vendor lane to let a different model family write the code when you need a second opinion.

🆔 @hackernewsgithubprojects

📦 layer6ai-labs/agentic-monte-carlo

Agentic Monte Carlo

Artificial intelligence agents can now simulate and optimize their own decision-making processes much like humans play out scenarios in their heads. The agentic-monte-carlo project introduces a way to run reinforcement learning simulations for black-box language agents by combining Monte Carlo methods with custom value functions. Instead of letting an agent guess blindly, this system generates multiple parallel paths, evaluates their potential, and dynamically resamples the best paths forward. By training a specialized model on past trajectories, it acts as a smart guide that steers the agent through complex, multi-step tasks like online shopping or scientific reasoning with far greater success.

🆔 @hackernewsgithubprojects

📦 sensengao/pixworld

PixWorld

Creating a fully explorable three-dimensional world from a text prompt or a single image normally takes a massive amount of computing time, but a new system can generate a complete 3D scene in just over half a second. This project, called pixworld, achieves this speed by combining both 3D scene generation and reconstruction into a single AI model. Instead of translating images through complex compressed layers, it works directly with pixels to output highly detailed 3D Gaussian spaces. It is roughly one thousand times faster than previous approaches, showing us that real-time 3D world building might finally be within our reach.

🆔 @hackernewsgithubprojects

📦 michaelshimeles/boring-computers

Boring Computers

Spin up custom, fully isolated Linux computers for your AI to control in just milliseconds. With boring-computers, you get actual virtual machines complete with a desktop, a browser, and preinstalled coding agents that run safely on your own hardware. You can ask an AI to build a game, and it will write the code, launch it, and hand you a live link to play. If the AI gets stuck, you can instantly fork the machine's exact state in milliseconds to test a different path. It is the ultimate playground for running untrusted code safely. Check out the repository to start hosting your own AI sandboxes today.

🆔 @hackernewsgithubprojects

📦 vita-epfl/everanimate

EverAnimate

Animate full-length, minute-scale human videos without watching your characters morph into totally different people or seeing the background dissolve into a blurry mess. Most video generators suffer from massive visual drift and identity loss when rendering longer clips because they forget the context from earlier frames. EverAnimate fixes this by using a clever latent memory system that anchors the generation process, alongside a smart restoration trick that adjusts the video's flow to prevent background flickering and quality loss. The best part is that it is fully open-source and light enough to fine-tune on standard hardware. Grab this tool to finally make long, consistent character animations that actually hold together from...

🆔 @hackernewsgithubprojects

📦 gkurt/tegaki

Tegaki

Turn any font into animated, self-drawing handwriting instantly without manually drawing a single line. This clever project processes font files behind the scenes by extracting glyph paths, converting them into clean digital skeletons, and automatically figuring out the natural stroke order and width. It means you can drop in your favorite font and get a beautiful, realistic writing animation that draws itself stroke by stroke. It supports all major web frameworks, works right in your terminal to output lightweight SVG files, and even features an interactive studio to customize your text. Check it out to bring your web projects to life.

🆔 @hackernewsgithubprojects

📦 yeje-cpu/talk-to-fengge

Talk to Fengge

Clone anyone's voice and personality to have highly realistic, real-time voice conversations with less than a second of engineering delay. By combining LiveKit for audio streaming, open-source voice cloning tools, and deep personality injection, this project lets you simulate complete characters instead of just generating robotic text-to-speech. You can feed it brief audio samples alongside speech patterns and catchphrases to turn an AI into a specific person you can actually talk to over a simulated phone call. Experience the future of natural, voice-cloned AI companions today by trying out the pre-built templates.

🆔 @hackernewsgithubprojects

📦 aiksa2090/agentic-swarm

Agentic Swarm

Imagine running five specialized AI agents that collaborate in an infinite brainstorming loop, entirely offline on your own computer. The project agentic-swarm lets you do exactly that by running Google Gemma-4 models locally. It solves the massive video memory problem of running multiple AI models at once by using a clever trick: it dynamically loads, runs, and completely clears each model from your graphics card before the next agent takes its turn. You get a beautiful dark-mode interface styled like WhatsApp with full real-time streaming, allowing you to inject ideas and steer the conversation. It is a brilliant way to brainstorm complex ideas locally without sharing your data with cloud...

🆔 @hackernewsgithubprojects

📦 androidmalware/openclaw_termux

openclaw_termux

Turn your Android phone into an autonomous AI assistant that you can control entirely through everyday chat apps like WhatsApp, Telegram, or Discord. This repository provides a simple, one-command script that installs a full Ubuntu environment inside Termux without requiring root access, setting up the OpenClaw gateway in just a few minutes. Once active, you can interact with your mobile device remotely by sending text messages to analyze files, trigger system commands, read incoming notifications, and capture photos directly from your phone. It is a brilliant way to recycle an old device into a private, always-on pocket server that acts on your commands.

🆔 @hackernewsgithubprojects

📦 zjunlp/lightmem2

LightMem2

LightMem2 is the lightweight runtime framework that finally stops context bloat from making your AI agents incredibly slow and expensive. When you run AI agents for a long time, their chat history and massive tool outputs quickly clog up the system, sending your token costs through the roof. This clever plug-and-play plugin acts like a smart filter, stabilizing your prompts and trimming away useless, oversized details before they can poison future turns. In real-world sessions, it slashes input tokens by over ninety-five percent and cuts hosting costs by eighty-seven percent. If you want faster, cheaper agents that do not lose their minds over time, this is the tool to install.

🆔 @hackernewsgithubprojects

📦 everyinc/hands-on-deck

hands-on-deck

hands-on-deck is the PowerPoint tool that lets AI agents edit and build slides using predictable JSON patches instead of fragile XML code. PowerPoint files are basically zipped XML, which makes it incredibly easy for an AI to corrupt a slide, miss brand guidelines, or mess up layout math. This tool solves that by letting agents write a simple, structured list of changes that run through a Python command-line interface. If an agent tries to modify a shape that does not exist, the tool rejects the entire patch, lists the actual shapes on the slide, and teaches the agent how to correct its own mistake instantly. It even includes an automatic...

🆔 @hackernewsgithubprojects