📦 tokenrhythm/neohorse

How NeoHorse Learns to Improve Itself

NeoHorse is the language model project that treats AI training as a self-improving loop. Instead of just building a smarter chatbot, the team created a system where the models actually do the work. They run tasks, record what goes right or wrong, and then use those results to update their own training data. It is like a student who grades their own homework and uses the corrections to study better. This creates a cycle where the system gets sharper at every step without needing a human to manually rewrite every lesson. The goal is not just a static tool, but a prototype that can keep getting better on its own.

🆔 @hackernewsgithubprojects

📦 yinminghuang/awesome-agentic-visual-generation-model

Awesome Agentic Visual Generation Model

Sort the chaos of visual AI by asking where a system truly takes the wheel. The awesome agentic visual generation model provides a clear ladder from fixed recipes to systems that learn from past work. It strips away hype to judge if a model actually changes its plan after seeing the result. This helps researchers spot genuine progress in creative tools. You finally get a fair way to measure real agency in visual generation.

🆔 @hackernewsgithubprojects

📦 nehanth/swarmllm

SwarmLLM: Run a 27B Model in Your Browser Tabs

A twenty-seven billion parameter AI model, one that normally requires a powerful server to run, is currently generating text by splitting its workload across a MacBook and an iPhone sitting in the same room. SwarmLLM is a project that takes a massive language model and slices it up so that your laptop, phone, and tablet each hold a small piece of the brain. No installation required, no cloud server, and no data leaves your house. It uses direct peer-to-peer connections between your browser tabs to pass tiny slivers of information back and forth, effectively turning your living room devices into a single, cooperative supercomputer.

📰 https://news.ycombinator.com/item?id=49602451

🆔 @hackernewsgithubprojects

📦 cpldcpu/pico-faces

A Face Generator on a One-Dollar Chip

Pico Faces is the one-dollar microcontroller project that generates human faces using a generative AI model. This tiny chip runs a latent flow diffusion transformer to create one hundred twenty-eight by one hundred twenty-eight pixel images of people in as little as five seconds. It solves the problem of running heavy AI on hardware with almost no memory by squeezing a complex model into just a few million parameters. The real surprise is that it can display these faces on a simple monitor or stream them over USB directly from the chip. You get a fully working AI image generator on hardware that costs less than a coffee.

📰 https://news.ycombinator.com/item?id=49489464

🆔 @hackernewsgithubprojects

📦 a-musiclover/tutti

Tutti turns sound into sheet music

Tutti is a deep learning system that listens to an audio recording and writes out the sheet music for you. It converts what you hear into a text format called ABC notation, which represents notes, rhythms, and structure in a way that machines can read and edit. This solves a long-standing problem where turning a song into a playable score required a human expert. Because the output is symbolic rather than just a waveform, you can now feed the results into other software for analysis or editing. The researchers built this using fully synthesized data to ensure the model generalizes well.

🆔 @hackernewsgithubprojects

📦 chaolongy/xemo-talker

Xemo-Talker Makes Faces Feel Real

Xemo-Talker stops making talking heads look robotic by explicitly controlling the emotions in the face. Usually, software just moves the mouth to match the audio, but this system looks at the voice to figure out if someone is sad, happy, or angry. It then changes the eyebrows, eyes, and head movement to match that specific feeling. This makes the final video look much more natural and human. It is built on top of existing face tracking tools but adds a smart layer for emotion. The result is a portrait that actually feels like it has a personality. This is a big step forward for realistic video synthesis.

🆔 @hackernewsgithubprojects

📦 jonnyoo/glance

Glance: Face Unlock for Your Mac

Glance is the Mac app that unlocks your computer with your face, just like Face ID on your iPhone. Instead of typing a password or pressing a physical button, a small animation appears in your laptop's notch and recognizes you instantly. It uses on-device AI to scan your face, ensuring your private data never leaves your computer. While it’s not as secure as a 3D sensor, it’s a clever convenience for quick access. If you want to skip the hassle of typing, this open-source tool offers a slick, modern way to get back to work fast.

📰 https://news.ycombinator.com/item?id=49623901

🆔 @hackernewsgithubprojects

📦 beescuit/chatflare

Chat with Friends Using Cloudflare Cache Hits

Chatflare lets you swap messages with a friend by flipping bits on Cloudflare cache hits. It solves the boring problem of standard chat apps by turning ordinary website requests into a communication channel. You send data by requesting specific file paths, which warms up the Cloudflare cache. When your friend checks those same paths, they get a cache hit, proving you sent a bit. This turns a normal web request into a one-bit signal. It is a clever proof of concept that shows you can talk over trusted domains without special servers. The result is a terminal app where you type a message and it gets encoded into HTTP requests.

📰 https://news.ycombinator.com/item?id=49397875

🆔 @hackernewsgithubprojects

📦 zhoushoujianwork/easyeda-agent

EasyEDA Agent: The AI Hands for Your Circuit Boards

EasyEDA Agent gives artificial intelligence the physical hands it needs to actually build your circuit boards. Instead of just suggesting code or text, it directly manipulates the schematic editor by placing real components and routing wires on the canvas. The coolest part is its circuit block library, which stores proven designs like USB interfaces so the AI can copy working circuits instantly rather than guessing every connection. It turns messy manual work into a clean, automated process where the machine places chips, runs checks, and fixes layout errors for you. You stop drawing every line by hand and start directing a digital assistant that handles the tedious details.

🆔 @hackernewsgithubprojects

📦 anbeime/skill

skill: Your AI Agent's Tool Library

skill turns your AI assistant into a specialist by giving it ready-made packages of instructions for specific jobs. Instead of teaching your chatbot how to handle PDFs or write code from scratch, you simply install a pre-packaged skill. This repository collects thousands of these ready-to-use modules, ranging from document processing to video creation, so your agent can tackle complex tasks instantly. It acts as a massive, organized library where you can browse, sort, and grab exactly the capability you need without writing a single line of setup code. It is the fastest way to make your AI assistant feel like a seasoned professional.

🆔 @hackernewsgithubprojects

📦 perseval-blr/dlss5-neuralscreen

NeuralScreen Makes Your Windows Desktop Look Like a Game

NeuralScreen runs the same neural rendering engine inside your Windows desktop as the latest games, making every window, photo, and video look sharper in real time. It does not just upscale your monitor; it uses a specialized chip to clean up and enhance the image right after your screen draws it, turning your entire desktop into a high-end visual canvas. The program is entirely portable, meaning you can drop it into a folder and run it without installing anything on your computer. It works with modern graphics cards and even lets you see the difference with a simple before-and-after slider.

🆔 @hackernewsgithubprojects

📦 opendcai/gamefactory-3a

GameFactory 3A: The AI Game Studio

GameFactory 3A is the open-source framework that turns a text prompt into a playable 3A game. It acts as a bridge between your ideas and major game engines like Unity, Godot, and Unreal. You tell a coding agent what you want, and it handles the heavy lifting. It generates 3D models, animations, audio, and even cinematic cutscenes. Then it writes the actual game code to make everything work together. The coolest part is that it runs locally on your own machine. This means you have total control over the creative process without renting expensive cloud servers. It’s like having a full game development team in your pocket.

🆔 @hackernewsgithubprojects

📦 kaybi-gh/k7

K7: Your Personal Netflix for Friends

Serve your own movie and music library to just your closest circle without handing your data to a giant tech company. K7 is a self-hosted media server designed specifically for small groups, giving you full control over who can access your content. The standout feature here is synchronized playback, which lets everyone watch the same show at the exact same moment, keeping the group in step. It works across phones, tablets, and desktops, making it feel like a private streaming service built for your household. Stop relying on public platforms that track every click and share your files with the world.

📰 https://news.ycombinator.com/item?id=49625127

🆔 @hackernewsgithubprojects

📦 ramonguthrie/comfyui-rbg-smartseedvariance

ComfyUI Smart Seed Variance Explained

Break free from identical AI images with ComfyUI Smart Seed Variance. This node solves the frustrating problem where changing your seed yields nearly identical results, especially in turbo workflows. It works by injecting smart, mathematically structured noise directly into your text embeddings during the early steps of generation. You get real artistic diversity—like shifting camera angles or changing lighting—while keeping your core prompt perfectly intact. It even includes specific settings to protect key words from getting lost in the noise. If you are tired of staring at the exact same composition over and over, this is the tool you need.

🆔 @hackernewsgithubprojects

📦 scottstevenson/superastra

SuperAstra: Play and Modify SNES Games with AI

SuperAstra lets you talk to a running Super Nintendo game and change it in real time. You simply type what you want, like giving yourself a cape every time you collect a coin, and the artificial intelligence investigates the game's memory to make it happen. It reads the screen, finds the right code, and applies the change between frames so the game never stops. If it makes a mistake, you can easily undo it and restore your progress. This is fascinating because it turns a classic console into a living lab where you can test ideas instantly.

📰 https://news.ycombinator.com/item?id=49625931

🆔 @hackernewsgithubprojects

📦 getopenpost/openpost

One draft, every platform: OpenPost

Turn one idea into ready-to-post content for every social network you manage. OpenPost is a self-hosted workspace where you write a single draft, then tweak the text and timing for each channel before publishing. It solves the tedious task of rewriting the same update for LinkedIn, X, and Mastodon, ensuring your voice stays consistent everywhere. The standout feature is its shared draft system, which lets you adjust media and scheduling for each destination while keeping a master version for your records. You can even run the whole stack on your own server with Docker. If you want to stop copying and pasting between apps, this is your new home base.

📰 https://news.ycombinator.com/item?id=49626154

🆔 @hackernewsgithubprojects

📦 kamaienterprises/aec-geometric-bench

How Good Are AI Models At Reading Blueprints?

AEC Geometric Bench lets you see exactly how well artificial intelligence actually reads construction blueprints. Most models can spot a room's general shape, but they fail to identify specific doors, windows, or fixtures needed to price a build. This project tests six systems on real architectural sheets using the same strict rules. The results are surprising: the specialized tool reads the drawings accurately, while the biggest general models miss most of the critical details. If you are building tools for construction, this open dataset gives you the real numbers to back up your claims, proving that general intelligence alone is not enough to handle precise technical documents.

🆔 @hackernewsgithubprojects

📦 tigerless-labs/agent-memory

Agent-Memory: AI That Actually Remembers

AI coding tools usually forget everything the moment you close the session, but agent-memory fixes that by keeping your data as simple text files instead of hidden databases. This means you can always open them and read exactly what your AI learned, without needing any special software or online accounts. It works locally on your computer, so different tools can share the same memory instantly. Because the data is just plain words, you can delete the search index at any time without losing your notes, and you can even edit the memories yourself with a standard text editor.

🆔 @hackernewsgithubprojects

📦 innate-labs/noobi.ai

Noobi.ai: The Desktop Agent That Codes Your Games

Noobi.ai is the local-first desktop agent that actually turns a simple text prompt into a fully playable browser game. It’s not just a code generator; it’s a complete development loop running right on your machine. You type what you want, and the agent builds, reviews, and compiles the game locally using a specialized app server. The best part? It’s private. Your prompts and code never leave your computer, which is a massive win for anyone worried about data privacy. It takes the friction out of coding by handling the heavy lifting of syntax and logic so you can just play.

🆔 @hackernewsgithubprojects

📦 primefoldtools/andon

Andon: Stop AI Mistakes From Repeating

Andon is the quality control system that actually stops your AI from repeating the same errors. Most tools just try to make the output sound better, but Andon borrows a hard lesson from factory floors to make sure a mistake never happens twice. If your agent claims a task is finished without actually testing it, or forgets what you agreed on last week, this is the fix. It uses a simple hook to catch those false claims and logs the defect so the system learns from it. It is not just a memory file; it is a discipline that turns every mistake into a permanent guard.

🆔 @hackernewsgithubprojects