📦 gen-verse/recuris

Recuris

Boost your long-horizon AI agents by evolving their memories instead of updating their weights or rewriting their prompts. This clever framework pairs a frozen agent with a state-grounded skill memory, capturing real execution traces to pinpoint exact points of failure. Instead of modifying a massive, bloated prompt, a meta-agent target-patches only the broken memory component, which must then pass a strict, model-free validation gate to survive. It is completely training-free and model-agnostic, meaning a memory evolved on one model transfers to others seamlessly. It is a brilliant way to slash agent errors on complex, multi-step tasks.

🆔 @hackernewsgithubprojects

📦 zhangzjn/awesome-smart-glasses

awesome-smart-glasses

Smart glasses are evolving from simple head-mounted displays into fully proactive, first-person intelligence platforms that can map what you see and intend directly into real-world actions. The awesome-smart-glasses repository acts as a master map for this new era, compiling cutting-edge research, hardware profiles, and real-world datasets. Instead of just listing gadgets, it tracks critical foundational capabilities like spatial mapping, first-person perception, and secure personal memory across diverse industries from healthcare to robotics. It is an invaluable goldmine for anyone wanting to see how wearable AI will actually integrate into our daily lives. Check it out to see where wearable tech is heading next.

🆔 @hackernewsgithubprojects

📦 atongwang/world-model-simulators

world-model-simulators

Artificial intelligence is trying to build entire virtual realities from scratch, but nobody knows if they actually obey the laws of physics. That is why the world model simulators repository is such a goldmine. It gathers two hundred ground-breaking papers tracking how close artificial intelligence is to becoming a true physics simulator. Instead of just generating pretty videos, it categorizes research on how artificial worlds handle real gravity, collision, and object interaction. If you want to see how robots are training in purely hallucinated environments without falling through the floor, you need to check this out.

🆔 @hackernewsgithubprojects

📦 leaplabthu/opdvr

Improve LLM Reasoning with OPDVR

Supercharge your AI model's math and reasoning skills by aligning teacher guidance with factual correctness. Traditional model distillation transfers knowledge from a smart teacher to a student, but it relies on dense token-level feedback that is completely blind to whether the actual final answer is correct or not. This project, opdvr, fixes this with an elegant, zero-hyperparameter gate. If a student's reasoning path leads to a wrong answer, the system instantly masks out conflicting teacher guidance. It basically ensures the student only learns positive lessons from correct steps, leading to massive accuracy gains across tough math benchmarks without any extra tuning.

🆔 @hackernewsgithubprojects

📦 dixin-lab/cft

cft

cft is an image relighting tool that keeps your original photo details perfectly intact while completely changing the lighting. Standard AI image editors often warp face shapes, change backgrounds, or lose the original person's identity when shifting from bright sunlight to moody shadows because they treat the process as entirely new image generation. Instead of starting from scratch, this project solves the problem by modeling the transition as a smooth path that directly carries over existing features. This means the AI isolates and shifts only the light rays while locking in your underlying facial geometry, structure, and identity. You get realistic, complex lighting adjustments that look completely natural every time.

🆔 @hackernewsgithubprojects

📦 emwstudio/voxemw

voxemw

Have a voice chat on your browser with a virtual anime companion using your own cloned voice, responding in under one and a half seconds. voxemw brings a complete virtual companion pipeline to your Apple Silicon Mac, managing voice activity detection, speech-to-text, and local voice cloning while utilizing a cloud brain for low-cost conversational intelligence. The standout feature is its real-time 3D VRM avatar rendering with server-side mouth shapes perfectly synced to every spoken word, reacting dynamically with natural expressions and head movements. It is an amazing way to experience ultra-low latency, interactive digital humans running directly on your local hardware.

🆔 @hackernewsgithubprojects

📦 zhoushen1/qurec

Qurec

An academic project called Qurec is solving the headache of ruined photos by fixing multiple types of image damage all at once. Instead of needing one tool for blur, another for noise, and a third for bad lighting, this clever system uses query-guided instructions to clean up your pictures in one go. It works by guiding the restoration process with specific queries and calibrating how the model looks at both tiny details and the bigger picture. It is a smart, unified approach to image repair that actually understands what needs fixing, making photo restoration a whole lot smarter.

🆔 @hackernewsgithubprojects

📦 teorth/erdosproblems

Erdos Problems

Erdosproblems is the community database that brings order to the legendary, unsolved mathematical challenges left behind by Paul Erdos. Instead of letting these brilliant puzzles sit scattered across textbooks, mathematicians Terence Tao and Thomas Bloom are crowdsourcing a central hub to track them all. The project links each problem directly to known integer sequences on the Online Encyclopedia of Integer Sequences and even tracks whether a solution has been verified by computer proof assistants like Lean. It is a playground for anyone who loves math, letting you help calculate sequences or track down historical proofs. Check it out to see if you can solve a piece of math history.

📰 https://news.ycombinator.com/item?id=49461053

🆔 @hackernewsgithubprojects

📦 joseairosa/recall

Recall

Recall is the model context protocol memory server that finally solves the problem of AI session amnesia. Instead of losing your hard-earned context, setup rules, and architecture choices every time you restart a chat or hit a token limit, this tool acts as a persistent brain extension for Claude. It stores critical patterns, decisions, and preferences once, then automatically pulls them in across any future conversation. You can self-host it using Redis or Valkey to keep your data fully private, or use their simple cloud service. It is basically the ultimate way to stop repeating your project requirements to AI assistants over and over.

🆔 @hackernewsgithubprojects

📦 starseed12345/queststack

Queststack

Queststack is the desktop application that finally gives your original Meta Quest headset complete freedom from corporate software limits. Instead of letting your old virtual reality hardware collect dust in a locked ecosystem, this clever tool automates a complex chain of security exploits to grant you full root access and an unlocked bootloader. It safely copies your headset's system partitions to your computer, temporarily slips a older vulnerable bootloader into place, unlocks the hardware, and safely restores everything back to normal. It is an incredibly safe, automated way to rescue your hardware, extend its life, and truly own the device you paid for.

📰 https://news.ycombinator.com/item?id=49461347

🆔 @hackernewsgithubprojects

📦 thu-keg/ruverbench

RuVerBench

RuVerBench is the rubric verification benchmark that finally measures how reliably AI judges check complex, long-form developer work. If you have ever asked a language model to grade a sprawling research report or evaluate a complicated coding run, you know it is incredibly difficult to tell if the AI judge's feedback is actually accurate. This tool solves that by matching AI grades against strict human-verified gold labels across thousands of test cases. It is an incredibly clever way to benchmark meta-evaluation strategies, meaning you can finally stop guessing if your AI grader is doing its job right or just hallucinating its feedback.

🆔 @hackernewsgithubprojects

📦 justinwetch/higagentskills

HIG Agent Skills

Give your favorite AI agent the power to design flawless apps that look and feel exactly like native Apple software without burning through your context budget. The higagentskills repository condenses the massive, text-heavy Apple Human Interface Guidelines into a set of hyper-focused, information-dense reference files. Instead of wasting thousands of tokens on fluff, it uses a clever keyword-routing system to feed your agent only the exact rules, measurements, and design principles it needs for a specific prompt. It is like giving your AI assistant a pocket cheat sheet of Apple's secret design playbook so it can build perfectly styled apps on the first try.

🆔 @hackernewsgithubprojects

📦 alpindale/gpt2.cmake

GPT-2 in CMake

Run the GPT-Two language model inside your build system configuration tool using pure CMake scripts. Yes, someone actually built a fully functioning neural network runner using nothing but the tool developers normally use just to compile C++ code. It bypasses heavy machine learning frameworks entirely, executing the model's math with fixed-point integer arithmetic during your build setup. You just download the standard model files, run a preparation script, and let your compiler configuration generate text. It is a completely absurd, brilliantly useless hack that shows just how far you can push a basic build tool when you are bored.

📰 https://news.ycombinator.com/item?id=49412909

🆔 @hackernewsgithubprojects

📦 jetbrains/go-modern-guidelines

go-modern-guidelines

The guidelines extension that keeps AI coding agents from generating outdated Go code. When you ask models to write Go, they constantly fallback to old habits, like writing nested loops instead of using modern slice utilities, mostly because their training data is stuffed with legacy patterns. This clever project from JetBrains solves that by injecting explicit rules directly into tools like Claude Code and Cursor. It forces your AI to swap out ancient boilerplate for clean, native features from recent Go releases. Now your agent automatically writes modern code on the very first try, meaning you have way less junk to refactor later.

🆔 @hackernewsgithubprojects

📦 yashmulgaonkar/flightscnr_pi

FlightScnr Pi

FlightScnr Pi is the hardware tracking system that turns a Raspberry Pi and a round touch display into a beautiful, dedicated radar station for your desk. By combining live aircraft data with real-time marine traffic, this project maps out everything moving in the sky and on the water right around you. You can swipe through detailed flight paths, check weather conditions, overlay local wildfire maps, and even stream live air traffic control audio straight to a speaker. Everything is controlled through a local web portal, giving you a custom, interactive radar without typing a single terminal command.

🆔 @hackernewsgithubprojects

📦 jdilla1277/agentcad

AgentCAD

Your coding agent can now design physical three dimensional models directly from your workspace. AgentCAD is a local command line tool and Model Context Protocol server that gives artificial intelligence coding assistants the ability to execute Python scripts, export step files, and generate high-quality three dimensional renders. It automatically provides interactive browser previews and side-by-side volume comparisons so you can visually verify every design iteration. It runs entirely locally with zero configuration required, translating complex geometric metrics into clean structured data that your developer agent can easily parse. Just install it and let your agent handle the math while you watch your physical designs come to life.

📰 https://news.ycombinator.com/item?id=48381402

🆔 @hackernewsgithubprojects

📦 johannhotzel/molecules

Molecules

You can turn any 3D mesh in Unity into a squishy, fully interactive soft body that explodes on command. A project called Molecules makes this possible by running a massive particle physics simulation entirely on your graphics card. Instead of relying on slow CPU calculations, it uses advanced math and GPU compute shaders to handle thousands of wobbling particles, custom collisions, and structural constraints in real time. It can even voxelize 3D models on the fly, turning solid shapes into bouncy, jelly-like structures you can poke and blast apart. It is a remarkably clever way to bring highly reactive, satisfying soft-body physics to your game engine.

🆔 @hackernewsgithubprojects

📦 miaai-lab/qwen3.8-flash-next-dual-dgx-sparks

Qwen3.8 Flash Next Dual DGX Sparks

Qwen3.8-flash-next-dual-dgx-sparks is the serving recipe that squeezes an insane three-fold increase in active memory out of powerful AI hardware. By squeezing the key-value cache down to a tiny four-bit format, this configuration lets you fit almost three million tokens into memory where you used to only get nine hundred thousand. It pushes context limits up to a massive one million tokens while keeping retrievals completely accurate up to one hundred and twenty-eight thousand. It is a clever way to run huge, fast models on high-end chips without running out of breathing room. Fire up this setup to supercharge your large-scale language model workloads today.

🆔 @hackernewsgithubprojects

📦 tolitius/cupel

Cupel: The Ultimate Local LLM Benchmarking Tool

Score local and cloud language models on your own machine using custom prompts and automated judges. Instead of guessing which open-source model actually works for your specific task, this tool lets you run structured, multi-turn evaluations using an automated grading rubric from zero to three. It auto-discovers local servers running on your machine, handles complex reasoning models by isolating their thinking blocks, and serves up a beautiful web dashboard. You get clear leaderboards tracking accuracy, speed, and cost side-by-side. It is the perfect way to stop guessing and start measuring. Search for cupel on GitHub to start grading.

🆔 @hackernewsgithubprojects

📦 get-concord-ai/concord-mcp

Concord MCP

Concord MCP is the open-source communication layer that finally lets different AI coding agents talk directly to each other inside your repository. If you are running multiple tools like Claude Code, Cursor, or Gemini at the same time, they normally operate in total isolation, leading to overwritten files and wasted effort. This clever local tool acts as a shared Model Context Protocol server that gives your agents a unified workspace. It allows them to claim files before editing, send live coordinate messages, and hand off tasks with full context and history. You get a clean, keyboard-driven dashboard to watch your AI team collaborate in real time without any messy conflicts.

📰 https://news.ycombinator.com/item?id=49464704

🆔 @hackernewsgithubprojects