📦 amazon-science/patientagentbench

PatientAgentBench: Testing Health AI

PatientAgentBench puts AI health assistants through rigorous, realistic role-play simulations to measure how well they handle real-world medical tasks. Instead of simple quizzes, the framework pits a smart chatbot against a simulated patient with a detailed health history. The AI must then navigate complex workflows like booking appointments, managing prescriptions, or recommending telehealth visits. Every interaction is judged by a panel of AI jurors using over a hundred criteria covering safety, helpfulness, and task completion. Because every scenario is generated on demand, there are no memorizable answer keys, ensuring the system truly understands healthcare rather than just guessing. This approach provides a honest, standardized way to compare how different large...

🆔 @hackernewsgithubprojects

📦 wloner0809/cast

How Cast Solves the Credit Assignment Problem in LLM Games

Cast turns game solvers into turn-level teachers for large language model agents, fixing the sparse reward problem by measuring how close each move gets the agent to victory. This method provides dense, per-action feedback that accelerates training and improves generalization across puzzles and real-world tasks. By using a deterministic solver to estimate remaining work, the system shapes rewards to credit progress, making it easier for agents to learn complex strategies without needing full outcome signals. The result is a faster, more reliable way to train AI that can solve Sokoban, Minesweeper, and Rush Hour while transferring skills to new environments. This approach bridges the gap between simple prompting and deep...

🆔 @hackernewsgithubprojects

📦 sqliteai/waste

WASTE: Running 2.78 Trillion Parameters on a Laptop

Run the full 2.78-trillion-parameter Kimi K3 model beyond available RAM by streaming activated weights directly from NVMe. A dependency-free, embeddable C inference engine. WASTE keeps the model trunk in memory, streams selected experts directly from disk, and uses the remaining RAM as a bounded expert cache. It keeps the model trunk in memory, streams selected experts directly from disk, and uses the remaining RAM as a bounded expert cache. Its current proof point is the complete open-weights Kimi K3 model: 2.78 trillion parameters, converted into a 982 GiB container and running on a 64 GB MacBook Pro at 0.49–0.54 tokens per second.

📰 https://news.ycombinator.com/item?id=49106591

🆔 @hackernewsgithubprojects

📦 subhan-code/amicro--micro-transitions-

Amicro: React Micro-Transitions Library

Amicro is the React component library that brings professional-grade micro-transitions and UI interactions to your frontend projects with zero configuration. Instead of manually coding complex animations or wrestling with CSS keyframes, developers can simply install it and use a command-line tool to add ready-made effects like 3D card carousels, arc layouts, and haptic-feedback buttons directly into their apps. The repository ships with over thirty-five pre-built, lightweight, and tree-shakeable components that work seamlessly with React and Tailwind CSS, ensuring your website feels polished and responsive without bloating the bundle size.

🆔 @hackernewsgithubprojects

📦 nxtg-ai/forge-orchestrator

Forge Orchestrator: Multi-AI Coding Coordination

Forge Orchestrator is a Rust tool that lets multiple AI coding assistants work on the same project without stepping on each other's toes. It stops tools like Claude and Codex from accidentally overwriting the same files by using file locks, so you never lose code to conflicting edits. It also captures decisions and patterns in a knowledge base that all tools can access, meaning the next session starts with context instead of a blank slate. You just generate a task plan from your specs, and the orchestrator coordinates the work across different AI providers. It’s basically a project manager for your bots, ensuring they actually collaborate instead of fighting.

🆔 @hackernewsgithubprojects

📦 factory-ai/cursed-plugins

Cursed Plugins: Code Reviews That Roast You

The cursed plugins repository from Factory AI transforms your codebase into a source of dark humor by using AI agents to generate absurd analysis reports. Instead of standard technical feedback, you can ask your coding assistant to write a dating profile for your repository, conduct a therapy session for your files, or produce a dead-serious migration plan from modern JavaScript back to COBOL. These modules work with popular AI coding tools to turn messy legacy code into entertainment, letting you share funny roasts of your own projects on social media. It is a lighthearted way to cope with technical debt without taking the stress too seriously.

📰 https://news.ycombinator.com/item?id=49124125

🆔 @hackernewsgithubprojects

📦 puffinsoft/peek-cli

Let AI Agents See Your Browser

Let coding agents see your browser. Peek-cli allows artificial intelligence assistants to capture a screenshot of any open tab in your browser. Your agents can now iterate on frontend designs until they are flawless. Works with Claude Code, Codex, Copilot and many more. See it in action: It works by using a browser extension to stream screenshots over web sockets. For security, you need to connect the agent once on every startup. Start the web socket daemon. Connect your browser. You are good to go. View available tabs. Capture screenshot. Install the chrome extension. Install the cli. Install the skill. For claude code. For codex. Is this safe?

📰 https://news.ycombinator.com/item?id=48799078

🆔 @hackernewsgithubprojects

📦 llamastash/llamastash

LlamaStash: Launch Local AI in Seconds

LlamaStash is the terminal tool that lets you launch local AI models with a simple init wizard and zero technical overhead. It solves the headache of complex setup by automatically detecting your computer's hardware, finding AI models you already downloaded, and starting a chat interface in just one step. Instead of wrestling with command-line arguments, you get a fast, interactive screen to manage your models easily. It works as a simple command, a visual app, or a background service, making private artificial intelligence actually accessible on your own machine. You finally get to run smart, local language models without needing to be a system expert.

📰 https://news.ycombinator.com/item?id=48369499

🆔 @hackernewsgithubprojects

📦 pachterlab/paperpush

PaperPush: Auto-Fill Journal Submissions

PaperPush automatically fills out academic journal submission forms using browser automation and optional artificial intelligence extraction. It supports dozens of major publishers and preprint servers by generating a local template file from your manuscript metadata. You can manually edit this file, use an AI agent to populate it, or let the tool read your paper directly. The software then opens the specific submission portal for your chosen venue, logs you in, and fills every field automatically. Crucially, it never clicks the final submit button, leaving you to review the completed form and press confirm yourself.

📰 https://news.ycombinator.com/item?id=49112424

🆔 @hackernewsgithubprojects

📦 jintae-00/redesign

ReDesign: Turn Flat Images Into Editable Designs

ReDesign is the AI agent that finally turns flat images back into editable design files. Most people lose their original source files and are stuck with uneditable flat exports, but ReDesign solves this by breaking a single image down into real text, vector shapes, and stacked layers. It works like a smart detective that looks at the picture, makes a guess about which tool to use, and then checks its own work before moving to the next piece. The most fascinating part is how it peels the image apart layer by layer until you have a clean, editable structure instead of just pixels.

🆔 @hackernewsgithubprojects

📦 tum-avs/chat2scenic

Chat2Scenic

Chat2Scenic transforms plain English descriptions into fully executable autonomous driving test scenarios in just a few steps. Instead of manually coding complex traffic simulations, this framework uses an interactive chatbot to interpret your natural language requests and iteratively refines them into precise Scenic script files. It leverages retrieval-augmented generation to ground the creation process in real-world regulatory documents and existing code examples, ensuring the output is both legally compliant and technically accurate. The system breaks the scenario down into logical components like vehicle behavior and environmental conditions, then assembles them into a ready-to-run script for the Carla simulator. This approach dramatically outperforms traditional methods, achieving a seventy-six percent compilation success...

🆔 @hackernewsgithubprojects

📦 skrabe/tweakcc-fixed

tweakcc-fixed: Customize Claude Code Prompts

Claude Code only exposes so much through its settings. tweakcc reaches the rest: it patches the installed binary directly — the cli.js, and the JavaScript baked into the native Bun build — so you can restyle the interface, rewrite the prompts Claude actually runs on, and change how it behaves. You pick what you want from a terminal UI, apply it in one command, and roll it back whenever you like. This is tweakcc-fixed, a fork of Piebald's tweakcc that keeps everything the original does and pushes further — over four times the prompt coverage, a deeper set of patches, and overrides that reach the native install where upstream stops.

🆔 @hackernewsgithubprojects

📦 aself101/agents-and-pipelines

Agents and Pipelines: AI Agents for Critical Thinking

Agents and Pipelines is a collection of fifty-seven specialized AI agents that treat software review as a debate rather than a checklist. Instead of just checking for syntax errors, this library equips you with thinkers like an Alien Frame Analyst who invents alternative problem definitions to expose hidden assumptions, or an Anxiety Reader that simulates a panicked operator to find weak spots before they break production. You simply copy these plain text definitions into your code editor, and suddenly your AI assistant doesn't just write code, it argues with it. It challenges your architecture using philosophical lenses and stress-tests your logic against perverse outcomes that no standard linter would catch.

🆔 @hackernewsgithubprojects

📦 intellisensing/self-in-space

Self-in-Space: Teaching Drones Spatial Awareness

Self-in-Space teaches drone AI to truly understand its own position and movement in 3D space. Most video models just see static pictures, but this project adds a special motion tracker that calculates how pixels shift between frames. By combining these movement signals with visual details, the AI builds a real sense of spatial awareness rather than just recognizing objects. It creates a benchmark to test how well drones can track themselves and reason about their surroundings. This is a game changer for making autonomous flying smarter and safer. Check out the repository to see how motion data transforms basic video understanding into embodied intelligence.

🆔 @hackernewsgithubprojects

📦 yc-software/qm

QM Multiplayer Agent Harness

qm lets teams give every employee their own isolated AI workspace that plays nice in Slack and the web. Instead of a single shared assistant, each person gets a personal sandbox with private memory, files, and tools that never mix with others. You can collaborate in group channels, run background tasks on a schedule, or build internal web apps together. It supports models like Claude and Codex, so you are not locked into one vendor. The best part is security: every action is audited, and admins can set approval rules to keep things safe. It is built for startups that need both independence and teamwork without the usual complexity.

📰 https://news.ycombinator.com/item?id=49126604

🆔 @hackernewsgithubprojects

📦 zwyang6/sage

SaGe: Teaching AI to Think Like a Detective

SaGe transforms how artificial intelligence understands images by forcing models to build mental maps of relationships rather than just spotting isolated objects. Instead of simply listing items it sees, this project uses a technique called scene graph thinking to trace how things connect in space and depth. By training the model to explore these structured connections step by step, it solves complex visual puzzles with remarkable accuracy. It is like showing the computer a map of the scene so it stops guessing and starts reasoning. This approach delivers state-of-the-art results on major benchmarks, proving that structured visual understanding is the key to truly smart machine perception. You can find the...

🆔 @hackernewsgithubprojects

📦 baihuajun24/spork

SPORK: Overlap Tool Latency with LLM Thinking

SPORK runs a lightweight probe alongside a large language model to predict and dispatch tool calls while the main model is still thinking. By guessing the next action early, it hides tool execution time behind the model's reasoning process. This overlap trick cuts tail latency significantly without changing the model's behavior, making agentic workflows noticeably faster when using think-mode architectures like Qwen.

🆔 @hackernewsgithubprojects

📦 bashalarmistalt/decimen-optical-transfer

Decimen Optical Transfer: Send Files with Light

Decimen optical transfer is the screen-to-camera tool that turns your monitor into a file server using nothing but light. It takes a file, shreds it into tiny data blocks, and then mixes them into a never-ending stream of animated QR codes displayed on the laptop. You open the receiving page on your phone, point the camera at the flickering screen, and watch the app rebuild the original file from the stream. The magic is in the fountain coding: the receiver captures any frames it can read, even if it misses some or grabs them out of order, and still perfectly reconstructs the data.

📰 https://news.ycombinator.com/item?id=49122553

🆔 @hackernewsgithubprojects

📦 lerd-env/lerd

Lerd: Rootless PHP Dev with Podman

Lerd lets you spin up local PHP development environments without Docker or root access by using rootless Podman containers. It automatically assigns your projects.test domains with instant HTTPS and isolates PHP and Node versions per project. The coolest feature is the built-in web dashboard that lets you manage sites, databases, and even debug code without touching the terminal. It is perfect for Linux and macOS developers who want a clean, frictionless setup. Check it out to simplify your local dev workflow.

📰 https://news.ycombinator.com/item?id=49127108

🆔 @hackernewsgithubprojects

📦 4rh1t3ct0r7/doom-regex

Doom Running on Regex

Doom runs entirely inside a single ninety-six megabyte text string using only fifty-four find-and-replace rules. Imagine the entire game engine, memory, and graphics stored as plain characters while a simple loop scans for patterns and swaps them out. Each swap updates registers, moves data, or draws a pixel until the next frame appears. It is a complete computer built from text substitutions, proving that even a classic shooter can be simulated without any actual arithmetic or code. The result is a byte-perfect match to the original game, rendered slowly but flawlessly. It is a fascinating look at how computation can hide in plain sight.

📰 https://news.ycombinator.com/item?id=49061275

🆔 @hackernewsgithubprojects