📦 ferrumox/rabbit

How Rabbit Runs a 2.8 Trillion Parameter Model on a Laptop

How Rabbit Runs a 2.8 Trillion Parameter Model on a Laptop

rabbit is a pure Rust engine that lets you run massive frontier AI models, like the 2.8 trillion parameter Kimi K3, directly on a standard consumer computer with just 128 gigabytes of RAM. Instead of trying to fit the entire model into memory, it keeps the essential parts resident in RAM while streaming the specific neural network experts onto the processor only when they are needed. This clever approach of loading thousands of tiny sub-networks from disk on demand means you can run some of the largest open-weight models available without needing expensive professional hardware.

📰 https://news.ycombinator.com/item?id=49093910

🆔 @hackernewsgithubprojects

📦 yu-fangxu/w2s-opd

Weak-to-Strong On-Policy Distillation

Make a strong AI smarter by teaching it from weaker models. That sounds backwards, right? Usually, you train a student using a teacher that is clearly better. But W2S-OPD flips the script. It takes a weak positive model and a weak negative model, compares their differences, and uses that contrast to build a fake, stronger teacher right inside the student's own math space. This lets the student improve without needing a massive, expensive expert model to guide it. It is a clever workaround for hitting the ceiling when you run out of bigger teachers to learn from. It is smart, efficient, and totally changes how we think about training.

🆔 @hackernewsgithubprojects

📦 h-embodvis/road

ROAD: 3D Shape Generation That Actually Makes Sense

Road is the 3D generation project that finally teaches computers to understand shape by aligning what they see with what they build. Most AI tools just guess geometry, but Road fixes that by locking a detailed 3D scanner onto the generation process. It matches every single point on the new model against a known perfect reference until they perfectly align. This means instead of blurry blobs, you get crisp, accurate textures that actually match the prompt. It is a clever way to force quality through strict point matching. Go check out the code if you want to build better 3D models.

🆔 @hackernewsgithubprojects

📦 anxmuy/agenticasr

AgenticASR: Fixing Speech Recognition in Real Time

AgenticASR actually rewrites speech transcripts as they happen to fix mistakes. Most speech tools just type what they hear, even when the speaker changes their mind or stumbles. This project adds a smart layer that listens closely and corrects the text on the fly. It removes filler words, fixes self-corrections, and updates earlier parts of the sentence when new information arrives. It works with any speech-to-text engine and handles both English and Chinese. You can use it for one-time cleanup or live streaming. The code includes everything needed to train, run, and test this refining process. It turns messy spoken language into clean, accurate written text automatically.

🆔 @hackernewsgithubprojects

📦 597358816/distilled-rl

Distilled RL: Smarter AI Learning

Distilled Reinforcement Learning is the AI training framework that teaches smaller models by using a smarter teacher to fine-tune every single word choice. Standard reinforcement learning usually just looks at the final answer to decide if the model succeeded, which is a blunt tool. Distilled RL changes the game by having a larger, expert model whisper corrections at every step, guiding the student’s grammar and logic in real time rather than just grading the result. This selective coaching helps tiny models understand complex math and reasoning far better than standard methods. You get impressive accuracy boosts without needing massive computing power.

🆔 @hackernewsgithubprojects

📦 epfml/looped-moe

Looped MoE: Share Expert Work, Cut Compute Costs

looped moe lets you slash the training cost of large language models by reusing the same brain cells across multiple layers. instead of teaching every single layer from scratch, it ties them together so they share weights, cutting parameter counts by up to four times without losing smarts. it works with popular models like olmoe and qwen three, proving that recycling knowledge is a clever hack for efficient ai. check it out if you want to train faster and cheaper.

🆔 @hackernewsgithubprojects

📦 borkdude/cljbang.el

Write Clojure in Emacs

Run Clojure code directly inside Emacs without ever leaving the editor or spinning up external processes. Cljbang lets you write familiar Clojure syntax that instantly compiles into native Emacs Lisp and executes right where you are. You get the elegant data structures and macros you love while keeping every function fully accessible to Emacs itself. It feels like magic because the performance is nearly identical to writing pure Lisp. This tool bridges two worlds so you can use modern syntax for everyday tasks while staying firmly inside your favorite environment.

📰 https://news.ycombinator.com/item?id=49004149

🆔 @hackernewsgithubprojects

📦 m4vic/socratic

Socratic: The Senior Engineer Question Bank for AI

Write code without guessing by letting an AI ask itself hundreds of critical engineering questions before building anything. This project packages the specific line of questioning a senior developer uses to catch risks early, turning that knowledge into a portable prompt for AI coding agents. Instead of blindly generating code, the AI silently reviews the task across domains like security, data, and infrastructure, self-answering routine details and only asking you about high-level decisions you must authorize. It stops when the plan is solid, ensuring the AI respects engineering best practices rather than just completing a checklist.

🆔 @hackernewsgithubprojects

📦 nomadamas/k-skill

K-Skill: The Korean Agent Toolkit

K-Skill is the agent toolkit that finally teaches AI to handle everyday Korean life tasks. Most language models stumble over local bureaucracy, but this project provides a library of specialized skills that let an AI assistant search for emergency room beds, check official housing prices, or even book intercity bus tickets using local systems. It is a fascinating collection of instructions and code that transforms a generic chatbot into a practical helper for living in Korea. Instead of just chatting, the AI can actively look up school lunch menus, find nearby clinics, or check government subsidy statuses.

🆔 @hackernewsgithubprojects

📦 narcooo/inkos

InkOS: Your AI Story Architect

InkOS is the story creation AI agent that finally keeps your novel’s plot threads and characters in sync while you write. It solves the problem of long-form stories losing their way by acting as a persistent memory bank for your entire world, from character traits to past plot points. The real magic is its narrative forecast feature, which lets you preview multiple possible futures for your next chapter without changing a single word of your current draft, letting you compare outcomes before you commit. You can also use it to translate full books or generate interactive game scripts.

🆔 @hackernewsgithubprojects

📦 yanhua1010/self-media-content-workflow

Self Media Skills: AI Content Workflow

Self Media Skills is an AI tool suite that turns vague ideas into ready-to-publish posts across platforms. It builds a creative brief, tracks trends, writes platform-specific copy, and generates video storyboards. The system refuses to auto-post, forcing five human approval gates like title and final copy checks before generating drafts. Every claim requires a source, and you can export finished packages for X or Xiaohongshu. It keeps your account safe by never logging in automatically, making it a transparent way to organize social media production without risking your login session.

🆔 @hackernewsgithubprojects

📦 xikhar/persona

Persona: Your Desktop’s New Expressive Avatar

Persona is the cross-platform desktop app that turns your voice assistants into expressive, animated characters right on your screen. Instead of just hearing text-to-speech in the background, you get a living, breathing avatar that reacts to what it is saying. It listens to the audio your computer is playing, detects when speech starts and stops, and instantly triggers natural lip-syncing and idle animations. You can even teach it new gestures to match the mood of the conversation. It works quietly in the background on Windows, macOS, and Linux, keeping your workflow uncluttered while adding a genuine human touch to your digital interactions.

🆔 @hackernewsgithubprojects

📦 ioai-official/ioai-2025

IOAI 2025: The AI Olympiad That Actually Counts Chickens

IOAI 2025 is the global AI competition repository that turns complex machine learning problems into tangible real-world challenges like counting chickens and authenticating antique paintings. Hosted in Beijing under UNESCO patronage, this archive collects the actual tasks, datasets, and reference solutions from the second International Olympiad in Artificial Intelligence. Instead of abstract benchmarks, participants tackled quirky, specific scenarios such as radar detection, restroom icon matching, and pixel efficiency. The project offers a transparent look into how top AI minds solve constrained, practical problems using Python, complete with detailed environment setups and official results.

🆔 @hackernewsgithubprojects

📦 richlegrand/bitbang-cli

Bitbang CLI: Remote Access Without Port Forwarding

Bitbang CLI is the remote access tool that works from behind any network with zero port forwarding or accounts. Instead of complex VPNs or SSH keys, you run one command on the remote machine and get a simple URL. Open that link in any browser to instantly get a full terminal, file manager, and network proxy. It even lets you pair devices using a six-digit code you read aloud over the phone. The connection is private and peer-to-peer, meaning the server only helps you connect but never sees your data. It is a clever, no-nonsense way to reach any computer without touching router settings or installing heavy software.

📰 https://news.ycombinator.com/item?id=49123789

🆔 @hackernewsgithubprojects

📦 zenbu-labs/terminal-browser

terminal-browser

Open a full website right inside your command line using terminal-browser. This tool turns your terminal window into a real web browser by rendering pixel data directly from the chromium engine. It is genuinely cool because it lets you surf the web while keeping your coding agent and terminal tabs in the same window. You can even open remote sites over ssh and interact with them using your mouse and keyboard. It works by sending special graphics commands to supported terminals to display the page smoothly. This means you can preview web pages or use web tools without ever leaving your favorite command line interface.

🆔 @hackernewsgithubprojects

📦 qaml-ai/camelai

camelai: AI Coding Assistant on Cloudflare

camelai turns ChatGPT-style chat into a real coding assistant by running directly inside the browser on Cloudflare infrastructure. The surprise capability is that it executes code in your browser sandbox, letting the AI fix bugs, run scripts, and preview results instantly without needing a complex server setup. It solves the friction of local development environments by giving you a full coding playground that starts immediately when you open the page. This is useful because developers can prototype, debug, and test ideas without installing heavy tools or configuring complex networks.

🆔 @hackernewsgithubprojects

📦 alexiglad/xm

XM: Explorative Modeling

This project introduces Explorative Modeling, a clever twist on generative AI that stops models from averaging their guesses. Instead of just producing one output and hoping for the best, XM generates a handful of candidate results at each training step and only trains on the single best match. This forces the model to commit to a clear, distinct mode rather than blending everything into a blurry mess. It is a simple but powerful way to improve image and video generation by making predictions sharper and more decisive. If you want to see how exploring multiple possibilities can unlock better end-to-end generation, this is the place to look.

🆔 @hackernewsgithubprojects

📦 elumenotion/guideants

GuideAnts: Make AI Work That Stays

Stop letting your best AI prompts vanish into the chat void. GuideAnts turns fleeting conversations into permanent, searchable projects by keeping your files, notes, and chats in one organized workspace. You can build reusable guides that package your favorite workflows, then publish them or embed them in other apps so your team never has to guess how to do a task again. It even edits Office documents right inside the chat. It is the smartest way to keep your AI expertise from evaporating.

🆔 @hackernewsgithubprojects

📦 alexzh3/codex-orchestrator

Let AI Review AI with the Codex Orchestrator

The Codex Orchestrator lets one artificial intelligence agent supervise another to catch mistakes that solo coding tools often miss. Instead of just writing code, a larger model plans the work and then assigns specific implementation tasks to a second, faster agent. Once the work is done, the first agent independently reviews every line of code and runs verification checks to ensure nothing was broken. This two-agent setup uses the unique strengths of different models, combining careful planning with fast execution while maintaining a full, inspectable history of every step.

🆔 @hackernewsgithubprojects

📦 dramaticshape/dramaticshapevoxelmod

Turn Classic Pokemon Into a 3D Diorama

Dramatic Shape Voxel Mod is the rendering overhaul that turns the flat Pokémon world into a clickable 3D diorama with real shadows and reflections. It extrudes the terrain into actual geometry and casts real shadows using a depth buffer, so buildings block your view just like they would in real life. The coolest part is the water, which isn't just a flat texture but a field of tiny voxel columns that rise and fall as waves and reflect the sky, sun, and nearby buildings with perfect clarity. You can even fight Pokémon battles directly on the landscape, complete with a cinematic camera that looks over the shoulder of your mon.

🆔 @hackernewsgithubprojects