📦 allspark-research/iris

Iris: The AI That Knows When to Stop Searching

Iris is an open-weight model trained to actually think while searching the web. Most AI just guesses, but Iris decides exactly what to search, how to read the results, and crucially, when it has enough evidence to stop. It uses a clever trick called discard-all to clear its memory when conversations get too long, preventing it from getting lost in the details. This means it can tackle hard questions without running out of steam. The repo even includes the exact testing harness used to measure its performance. So if you want an AI that hunts for facts like a detective, not just a parrot, this is the one to watch.

📰 https://news.ycombinator.com/item?id=49686115

🆔 @hackernewsgithubprojects

📦 abhi245y/ps5lm

PS5LM Plus: AI on Your TV

PS5LM Plus is the home console app that runs large AI models directly on your television. Instead of sending data to a remote server, this project uses the PlayStation's own graphics card to generate text locally. The standout feature is a live dashboard on your TV that displays real-time generation speeds, power usage, and temperature readings. You can download models from a built-in store, connect to your home Wi-Fi, and chat with the AI from any browser on your network. It turns your gaming setup into a private, offline brain that handles everything from quick questions to complex coding tasks right on your living room screen.

🆔 @hackernewsgithubprojects

📦 tribgames/mixdog

Mixdog: 63% Fewer Tokens

Get the same coding results while using significantly fewer tokens. Mixdog is an open-source desktop app that acts as a smart coding assistant, matching the performance of top tools like Codex while consuming sixty-three percent fewer tokens. It works by using a tiny, local model to decide exactly how much effort each step needs, so it stops burning resources on simple tasks. You can use it with subscriptions you likely already own, like ChatGPT or Claude, and it even lets you mix different models to balance speed and quality. This makes powerful AI coding far cheaper and more accessible.

📰 https://news.ycombinator.com/item?id=49930226

🆔 @hackernewsgithubprojects

📦 toddrjones/datasets

The Ultimate Data Cheat Sheet

Todd Jones, an applied microeconomist, has quietly compiled over one hundred free datasets that cover the messy, human side of the economy, from minute-by-minute time diaries to individual-level car crash records. This isn't a vague list of links; it is a curated, categorized index that tells you exactly where to find public data on crime, education, migration, and even the cost of living across counties. For students and researchers, this solves the single biggest problem in social science: the hours wasted hunting down obscure government archives just to find one specific variable. If you are building a model or teaching a course, this is your starting point.

🆔 @hackernewsgithubprojects

📦 dreamer-toby/stepquant

StepQuant: The Smart Way to Shrink AI Memory

Squeeze massive AI models into tight memory without losing their minds. StepQuant solves a tricky problem for neural networks that remember things over time. Instead of blindly chopping off precision everywhere, it figures out exactly which parts of the model's memory actually matter. It gives extra room to the bits that carry big, long-lasting errors and trims the rest. This lets you run big models on smaller hardware while keeping answers accurate. It is like packing a suitcase where you only fold the clothes you truly need. You get a much smaller footprint without the usual quality hit. Check it out if you care about efficient AI.

🆔 @hackernewsgithubprojects

📦 ajncsystem/shadowofthecolossus-pc

Shadow of the Colossus PC Port

Run the PlayStation exclusive Shadow of the Colossus natively on your Windows PC. This is not a standard emulator; it is a specialized loader that runs the game code directly on your processor while translating the graphics to Vulkan. The surprising part is how it handles the audio. Because the game uses specific console sound drivers, the launcher automatically searches your computer for existing emulator data to copy the necessary sound modules, meaning you get full audio without hunting for loose files. It even finds your save files and downloadable content on its own.

🆔 @hackernewsgithubprojects

📦 timescale/pg_vectorsearch

pg_vectorsearch: Fast Vector Search for PostgreSQL

Timescale’s new database extension, pg_vectorsearch, processes a hundred million data points using just a quarter of the memory required by existing graph-based methods. It achieves this speed by converting complex numbers into compact codes with strict error limits, allowing the system to instantly rule out irrelevant results before checking the original data. This project solves the heavy memory cost of finding similar items in a standard database, making massive scale searches practical without buying expensive hardware. It is currently an alpha version for testing, but it proves that high-performance similarity search can now live right inside your existing database tables.

📰 https://news.ycombinator.com/item?id=50086165

🆔 @hackernewsgithubprojects

📦 tornikebolokadze1-cyber/awesome-ai-pulse-georgia

Ask Your AI About This

Stop scrolling through endless GitHub lists to find the right AI tool. With this new feature, you can now ask your AI assistant directly for recommendations from a hand-picked collection of nearly four hundred verified projects. Instead of guessing which framework works for your specific coding task, you simply type a question into your editor, and the system searches a curated database to give you precise, categorized answers in both English and Georgian. It connects your AI assistant to a living library of developer resources, saving you hours of research. Now, your coding partner knows exactly where to look when you need help building with modern AI agents.

🆔 @hackernewsgithubprojects

📦 keplsiv/folderforge

FolderForge: The One-Click Mac Folder Styler

FolderForge lets you apply a single custom design to every folder in a specific directory using a smart rule system. Instead of manually changing icons one by one, you set a condition, like matching every folder named Receipts, and the app updates them all at once. It shows you exactly what will change before anything happens, so you can review the results safely. The tool keeps your original icons backed up locally, meaning you can undo the entire batch if you change your mind. Everything runs entirely on your machine with no accounts or data collection.

📰 https://news.ycombinator.com/item?id=49230965

🆔 @hackernewsgithubprojects

📦 openworkflowdev/openworkflow

OpenWorkflow: Your Code Survives the Crash

OpenWorkflow ensures your multi-step processes finish even if your server crashes or restarts in the middle of the job. It works by saving every single step of your code to a database, acting like a save point in a video game. If your application dies, the framework doesn't start over from scratch. Instead, it reads that database, skips all the work you already finished, and picks up exactly where you left off. This means you can build long-running tasks that take hours to complete, knowing they will survive power outages and deployment updates. The result is a system where your background jobs are permanently reliable without needing complex external infrastructure.

🆔 @hackernewsgithubprojects

📦 tintinweb/pi-subagents

Pi Subagents: The AI Fleet View

Pi Subagents is the pi extension that turns your coding assistant into a fleet of parallel workers. Instead of just chatting with one bot, it lets you spawn multiple autonomous agents that run at the same time, each handling a different part of your codebase. It mirrors the familiar tool patterns from Claude Code, so if you know that ecosystem, this feels right at home. The real kicker is the live widget and fleet view, which lets you watch every agent work in real time and steer them mid-run if they go off track. It is a clean, native way to orchestrate complex tasks without losing control.

🆔 @hackernewsgithubprojects

📦 myllmbox/qwen38-flash-next-recipe

Qwen38 Flash Next Recipe

A single desktop computer now serves a massive one hundred twenty five billion parameter artificial intelligence model at speeds that would have required a server rack just a year ago. This recipe project turns one Nvidia DGX Spark box into a local powerhouse, running the Qwen38 Flash Next model without needing a data center. The coolest part is the dynamic drafting system, which acts like a smart spell checker that guesses a few words ahead of time and then checks them instantly, making the output stream up to forty percent faster than previous versions.

🆔 @hackernewsgithubprojects

📦 feralfirmware/tailtalk

TailTalk: Talk to Your Old Macs

TailTalk is the network toolkit that finally lets modern computers talk to classic Macintosh systems. It rebuilds the entire AppleTalk protocol from scratch, meaning your current laptop can directly share files with a PowerBook from the nineties without any special drivers. The most surprising part is that it turns a simple USB dongle into a bridge, letting you drag and drop files between a modern computer and an old Mac using a standard serial port. It also lets you print to ancient LaserWriters from your current system. If you have old Apple hardware sitting in a drawer, this is the missing piece to bring it back to life.

📰 https://news.ycombinator.com/item?id=49678423

🆔 @hackernewsgithubprojects

📦 halluton/mindful-claude

Mindful Claude: Breathe While You Wait

Mindful Claude turns every ten-second wait for an AI response into a guided breathing session right inside your terminal. When Claude starts working, a simple animation appears, counting breaths with you to lower stress and sharpen focus. It is a clever plugin that uses dead time for a quick nervous system workout, offering patterns like box breathing or the physiological sigh. The animation vanishes the moment the answer arrives, so you stay in flow. You can toggle it on or off, or pick your favorite style, all without leaving your editor. It is a small, smart way to make coding less stressful.

📰 https://news.ycombinator.com/item?id=47313480

🆔 @hackernewsgithubprojects

📦 datalab-to/indra

Indra: The Tiny OCR That Beats the Big Guns

Indra turns any PDF or image into clean, usable text in a fraction of a second. Forget the usual heavy-duty tools that crawl along; this lightweight model chews through forty-nine pages per second on your standard processor. It doesn't just find words; it maps out the entire page, spotting headings and tables so your notes stay perfectly organized. It is fast, small, and shockingly accurate, handling everything from legal contracts to scanned receipts. You can grab the results as simple text, markdown, or detailed data boxes, ready for your next project. If you have been stuck with slow, clunky scanners, this is your new best friend.

🆔 @hackernewsgithubprojects

📦 cloudflare/artifact-fs

ArtifactFS: Stop Waiting for Clones

ArtifactFS is the filesystem driver that mounts massive git repositories instantly by skipping the initial full clone. It is built specifically to make code accessible for AI agents, sandboxes, and containers where waiting for a traditional download is simply too slow. The tool mounts the directory structure immediately so you can list files and run basic commands while the actual file contents download in the background. It cleverly prioritizes code and manifest files, meaning your most important data is ready first. You get a working environment in seconds rather than waiting for a massive transfer to finish.

🆔 @hackernewsgithubprojects

📦 excalidraw/excalidraw-mcp

Excalidraw MCP: Live AI Whiteboards

Excalidraw MCP is the Model Context Protocol server that finally lets AI models sketch directly inside your chat. Instead of just describing a diagram in text, it streams hand-drawn shapes right into the conversation, animating them in real time so you can watch the architecture build itself piece by piece. It works with any client that supports MCP Apps, meaning you can plug it into Claude, ChatGPT, or VS Code to get interactive visuals without leaving your current workflow.

🆔 @hackernewsgithubprojects

📦 sableclient/sable

Sable: The Web-Based Matrix Client

Sable is the web-based chat app that makes Matrix feel like a real application rather than a website. If you have ever wanted to self-host your messaging without giving up the polish of a modern interface, this is it. It runs entirely in your browser or desktop shell, giving you full control over your data while keeping the experience fast and responsive. The standout feature here is the deep customization layer, which lets you tweak themes and layouts directly from the client, something most other options ignore. It is not just a simple wrapper; it is a complete, stable client that respects your privacy.

🆔 @hackernewsgithubprojects

📦 i2cjak/backplane

Backplane: Watch AI Build Circuit Boards

A coding agent is editing your physical circuit board layout in real-time, and you are watching the copper traces move on a screen. Backplane connects AI assistants directly to KiCad and FreeCAD, letting them modify your hardware design files while you observe every change instantly. This solves the problem of blind edits, where you used to read code to understand a design. Here, the board, schematic, and three-dimensional model redraw themselves as the agent works. You can compare versions and even model the physical plastic case in the same loop. It runs on your own computer, keeping your files and tools private.

🆔 @hackernewsgithubprojects

📦 retroportingtoolkit/psxrecomp

Turn a PS1 Disc into a Real Program

Take a PlayStation disc and turn it into a native program your computer runs directly. P S X R E C O M P translates the console's original machine code into C, which compiles into a standard executable file. It is not an emulator pretending to be hardware; it is the game itself, running as a real application on your modern processor. The result is a standalone build that loads instantly, skips the slow boot sequence, and supports widescreen resolution out of the box. The framework handles the heavy lifting of code generation, so developers can port titles with configuration rather than rebuilding the engine.

🆔 @hackernewsgithubprojects

📦 api7/aisix

aisix: One API for Every AI Model

aisix is the open source AI gateway that puts one simple interface in front of every major language model. It solves the headache of juggling different coding standards by translating your requests so you can talk to OpenAI, Claude, or Gemini using the exact same code. The real standout feature is semantic routing, which reads the meaning of your prompt and automatically sends it to the best suited model for that specific task. This happens behind a single, fast binary that also handles security and caching. If you are building applications that use multiple AIs, this tool lets you swap providers without rewriting your code.

🆔 @hackernewsgithubprojects