📦 smileformylove/xscientist

xscientist

This project transforms scientific research into a git-like protocol where experiments are completely diffable, forkable, and reviewable just like open-source code. Instead of simply generating a static paper, xscientist runs a continuous pipeline that treats every research attempt as a trackable commit, mapping out your entire scientific exploration as a branching tree of code, metrics, and failures. It even links written manuscript claims directly back to the exact code execution node that proved them. It is basically git for science, letting you branch off previous experiments to build a fully verifiable, automated tech tree.

🆔 @hackernewsgithubprojects

📦 ireliya/terralogic

Terralogic

Terralogic is the benchmarking framework that finally helps artificial intelligence make sense of Earth observation data by coordinating complex, multi-step search missions. Instead of just looking at a satellite image, this system acts like a coordinator for specialized remote sensing tools across optical, radar, and infrared sensors. It breaks down complex natural language requests, builds a logical execution plan, and runs detailed analysis like counting specific structures or tracking small infrared targets over time. If a sensor tool fails or returns bad data, the system automatically self-corrects and swaps in a compatible tool to finish the job, translating raw data into verified, plain-English answers.

🆔 @hackernewsgithubprojects

📦 metainfer/metainfer

MetaInfer

Build a custom, high-performance language model inference engine from scratch using an automated, agent-driven compiler approach. Instead of downloading a massive, bloated framework with endless configuration options, this project introduces a brilliant concept called Large Language Model as Compiler. It uses a structured knowledge graph and a multi-agent system of coders, reviewers, and testers to automatically assemble a lean, single-path inference engine tailored precisely to your specific model and hardware. This eliminates thousands of lines of dead code, resulting in massive speedups over traditional setups. It is a fascinating glimpse into the future of self-assembling, hyper-optimized AI infrastructure.

🆔 @hackernewsgithubprojects

📦 jetxu-llm/codex-deepseek-bridge

Codex DeepSeek Bridge

Redirect your OpenAI Codex app requests to DeepSeek using a clever local proxy that keeps your existing setup completely intact. This tool hooks directly into your configuration so you can use deepseek-pro and deepseek-flash while keeping your plugins, MCP servers, and custom workflows running exactly as they are. The best part is your API key never leaves your machine, and you get a beautiful offline dashboard to track your cache hits and latency. When you want to revert, a single restore command puts everything back to the official defaults without leaving a trace.

🆔 @hackernewsgithubprojects

📦 thudm/scale-cua

ScaleCUA

Teach your AI models to control any desktop operating system like a human. This project introduces a pipeline that generates thousands of realistic computer tasks, like browsing the web or editing spreadsheets, and builds automated judges to verify if the AI actually completed them correctly. Instead of wasting time on tasks that are too easy or too hard, it uses a smart sampling method to constantly feed the AI challenges right at the edge of its current capabilities. It is a highly efficient way to train digital agents, making them smarter and much faster at navigating real software.

🆔 @hackernewsgithubprojects

📦 starnodes2024/comfyui-starnodes-modelconverter

comfyui-starnodes-modelconverter

You can now squeeze massive diffusion models down to just a quarter of their original size without ruining your image quality. The comfyui-starnodes-modelconverter is a custom node package for ComfyUI that introduces advanced, profile-based quantization to intelligently compress heavy models like Flux, SDXL, and Wan. Instead of a basic blanket compression, this project analyzes your model layer by layer and uses smart blacklists to keep highly sensitive elements like embeddings and normalizations in high precision while safely shrinking everything else. It even lets you split all-in-one checkpoints into clean, separate components for easier management.

🆔 @hackernewsgithubprojects

📦 fiigii/ai-comp

ai-comp

The optimizing compiler that solves a complex coding challenge by translating high-level code into highly efficient parallel instructions. ai-comp was built to conquer Anthropic's performance engineering take-home exam, which asks developers to manually optimize a tree-traversal and hashing program running on a simulated virtual machine. Instead of spending hours tweaking the machine code by hand, this project automates the entire process. It takes a human-readable description of the program and intelligently schedules instructions to run simultaneously on multiple virtual execution engines. The coolest part is how it automatically eliminates redundant memory steps, ensuring the simulated processor never wastes a single clock cycle waiting for data that it already has. It...

📰 https://news.ycombinator.com/item?id=48911420

🆔 @hackernewsgithubprojects

📦 smithandrewjohn/kindalive

Kindalive

Model robot emotions through simulated body chemistry instead of simple programming labels. With kindalive, you type a plain text paragraph, and a smart interpreter translates your words into real-time chemical impulses like dopamine, cortisol, and adrenaline. A core engine constantly calculates how these simulated chemicals decay and interact, translating the balance into twelve facial muscle movements. This drives a retro LED dot-matrix face that naturally contorts to show realistic, blended emotions without hard-coded rules. You get a fascinating look at emergent behavior, complete with a responsive web dashboard, plus ready-to-use adapters to bring your physical robot face to life on real hardware.

📰 https://news.ycombinator.com/item?id=48878966

🆔 @hackernewsgithubprojects

📦 yempik-ai/cowork-os

cowork-os

AI agents work best when they do not leave their results trapped inside disposable chat threads. cowork-os is an open-source workspace layer that saves your teams decisions, process knowledge, and project state in plain Markdown files you actually own. Instead of relying on opaque memory systems, this system uses your existing local files to keep your virtual assistant grounded and fully accountable. It translates unwritten team processes into structured workflows and runs automated routines, like daily briefs and pipeline updates, directly on your disk. You get an inspectable company brain that your team can easily review, correct, and control.

🆔 @hackernewsgithubprojects

📦 tristan666666/agent-island

Agent Island

Agent Island is the desktop status companion that finally keeps you in the loop with your background Claude Code and Codex sessions. If you run long coding agent tasks, you do not have to stare at an open terminal anymore. This lightweight, local-first app sits in your macOS menu bar or Windows taskbar, silently monitoring your active runs. It spins its logo while they work, flashes red if there is an error, and plays an alarm the second it is your turn to respond. You get instant updates, cost tracking, and weekly productivity reports with zero telemetry and telemetry. Just sit back, relax, and let your code write itself.

🆔 @hackernewsgithubprojects

📦 yale-nlp/llm-metacognition

llm-metacognition

Large language models can actually monitor and regulate their own thinking process just like humans do. The llm-metacognition repository collects and organizes scientific research exploring this exact boundary, providing a complete map of how models assess their own knowledge and correct their errors. It acts as an open library for the latest techniques to measure confidence, detect when an AI is guessing, and prevent hallucinations. By organizing these breakthroughs into a structured guide, it helps developers build much more reliable AI systems. If you want to understand how artificial intelligence is learning to think about thinking, this repository is the perfect place to start.

🆔 @hackernewsgithubprojects

📦 quhongyu/lamem-vla

LaMem-VLA

Control robots with natural language and vision using a clever system that remembers past actions to make better decisions. Most modern robot brains struggle to connect what they see with what they did a moment ago, making them clumsy. This project introduces a dual memory system that keeps track of both visual cues and physical movements, helping robots perform tasks much more smoothly. While the creator is still polishing up the code for release, keeping an eye on this repository is a great way to see how the next generation of smart, adaptable robots will actually learn to navigate our messy world.

🆔 @hackernewsgithubprojects

📦 yeanroot/halo-wa

halo-wa

halo-wa is the robotics training wrapper that finally makes online reinforcement learning practical for complex physical agents. Under the hood, it connects advanced vision-language-action policies with simulated physical environments to train robots using a hybrid-attention world model. While standard setups struggle to coordinate simulation physics and deep model learning, this framework packages everything into ready-to-run tasks like hammering or ringing a bell. By utilizing pre-trained world models as reliable digital training grounds, it lets robots practice and refine their physical skills safely in simulation before taking on the real world. You can download the pre-packaged data and start training physical policies immediately.

🆔 @hackernewsgithubprojects

📦 zhinianboke/xianyu-auto-reply

xianyu-auto-reply

xianyu-auto-reply is the multi-account management assistant that finally turns your secondhand store into a hands-off, automated storefront. It uses web sockets and browser automation to handle everything from answering customer questions with smart AI replies to auto-delivering digital keys and giving ratings. You can run multiple accounts, auto-refresh your listings to keep them on top of search results, and manage your inventory with zero manual clicking. It essentially acts as a twenty-four-seven virtual employee, making sure you never miss a sale even while you sleep. Set up this Python-based automation toolkit today to let your store run itself.

🆔 @hackernewsgithubprojects

📦 apache/ossie

Apache Ossie

Apache Ossie is the open-source specification that standardizes how we share business definitions across different database tools and artificial intelligence platforms. Normally, different teams and software packages define key metrics like revenue or customer count using completely different logic, leading to frustrating calculation errors. This project solves that problem by providing a single, vendor-neutral file format that any software can read and write. It translates complex data relationships into a universal language, giving your entire technical stack a single source of truth. It is a fantastic way to keep your database logic consistent, reliable, and perfectly aligned.

🆔 @hackernewsgithubprojects

📦 raroque/boop-agent

Boop Agent

You can now run a private personal agent right inside iMessage that uses your existing Claude Code or ChatGPT subscription instead of racking up expensive developer API bills. Named boop-agent, this clever system acts as a secure dispatcher on your Mac, receiving your texts and spawning specialized, isolated sub-agents to handle your tasks. It features a brilliant memory consolidation pipeline that runs daily to resolve contradictions and merge duplicates, plus a local browser automation tool for tricky websites. Because it stages external actions as drafts for you to approve before sending, you always stay in complete control of your digital life.

🆔 @hackernewsgithubprojects

📦 michael-manning/e-paper-climate-logger

E-Paper Climate Logger

The e-paper-climate-logger is the open-source hardware project that packages a custom-built, ultra-low-power weather station into a gorgeous, pocket-sized device. Running on an ESP32-S3 chip, it monitors temperature and humidity, stores the data onto an external memory chip, and renders historical trend graphs on an always-on paper-like display. The clever design consumes so little energy in deep sleep that it can run for over a week on a tiny battery. It is a brilliant example of smart power optimization, waking up for just a fraction of a second to log data before instantly putting itself back to sleep.

📰 https://news.ycombinator.com/item?id=48919013

🆔 @hackernewsgithubprojects

📦 vercel-labs/agent-skills

Agent Skills

Agent Skills is the AI toolkit that finally gives coding assistants real-world superpowers. Instead of letting your AI agent guess how to optimize your code, this repository packages production-grade instructions and scripts directly for them. Built by Vercel, it covers everything from performance-tuning React components and auditing deployments to managing view transitions and writing better documentation. The coolest part is the claimable deployment skill, which lets your coding partner package your local app, push it live, and hand you a preview link along with ownership. It turns your AI helper from a basic code generator into an active deployment partner.

🆔 @hackernewsgithubprojects

📦 firaslatrech/reqlog

reqlog

Inspect every single HTTP request hitting your server live in your browser with zero configuration. This neat tool spins up a real-time visual dashboard on localhost with just three lines of code. It instantly streams your network traffic, letting you inspect headers, dissect payloads, highlight slow bottlenecks, and even replay captured requests with a single click. There are no heavy databases, config files, or external dependencies to manage. It is a lightweight, local-only developer companion that blocks itself in production automatically to keep you secure. Toss out those messy print statements and watch your server traffic flow live.

📰 https://news.ycombinator.com/item?id=47437103

🆔 @hackernewsgithubprojects

📦 google/gnm

Inside GNM Google's New Open Source 3D Human Face Generator

Google just open-sourced a highly detailed 3D statistical model of the human head that includes fully controllable internal anatomy like eyeballs, teeth, and a tongue. The project, called gnm, lets you generate incredibly realistic 3D faces using simple mathematical sliders. Instead of manually sculpting, you can use pre-trained semantic samplers to generate distinct facial shapes or expressions like happiness and surprise just by typing plain text. Because it supports major frameworks like PyTorch, JAX, and TensorFlow right out of the box, developers can easily integrate these expressive, custom faces into their own graphics pipelines, virtual avatars, or AI research tools.

📰 https://news.ycombinator.com/item?id=48919820

🆔 @hackernewsgithubprojects