📦 solarkyle/jspace

Inside jspace: Can LLMs Detect Their Own Errors?

Stop relying on basic confidence scores to catch artificial intelligence mistakes. A fascinating new project called jspace analyzes the internal signals of large language models to find out if they actually know when they are hallucinating. By looking deep inside the model's residual stream during a massive twenty-five thousand prompt campaign, researchers trained a tiny three-hundred kilobyte classifier that predicts wrong answers better than the model's own output confidence. The coolest part is that this internal signal successfully transfers across different datasets in the same task family. However, the project's preregistered tests also revealed where these detectors fail, showing they are task-specific rather than universal error monitors. Check out the...

🆔 @hackernewsgithubprojects

📦 sgshuman/tjs-meal-planner

tjs-meal-planner

Organize your weekly grocery run with a smart assistant that structures your entire menu around a single trip to Trader Joe's. Instead of managing a chaotic daily schedule, this clever tool focuses on a streamlined weekly plan of just one breakfast, one lunch, and two dinners. It automatically sorts your custom shopping list to match the physical walking layout of your local store, so you can breeze through the aisles without backtracking. The best part is the offline support, meaning your list works perfectly even when you lose cell service inside the store. It is the ultimate stress-free shopping companion.

🆔 @hackernewsgithubprojects

📦 mosamlife/wpmgr

WPMgr WordPress Fleet Management

You can now manage and secure an entire fleet of WordPress sites from a single dashboard running on your own server, with zero reliance on third-party cloud services. An open-source tool called wpmgr provides complete self-hosted control over your WordPress installations, handling everything from updates and uptime monitoring to security audits. Its standout feature is a high-performance backup system that uses smart streaming technology to handle massive databases and media libraries without overloading low-power servers. Best of all, it encrypts backups on the client side before they ever leave the site, keeping your data completely private. It is the ultimate dashboard for taking back control of your web infrastructure.

🆔 @hackernewsgithubprojects

📦 aaromanov1985/audio-cassette-simulation

Audio Cassette Simulation

Audio Cassette Simulation is the command line tool that accurately recreates the unique sound profiles of vintage cassette tapes. Using the audio processing power of ffmpeg, this repository applies realistic tape hiss, frequency limits, and pitch wobbles to digital audio files or live web streams. You can make modern music sound like it was recorded on a clean nineties TDK tape, a warm seventies Sony cassette, or even a gritty, highly degraded Soviet MK sixty bootleg. It is an incredibly clever way to instantly inject nostalgic analog warmth and retro character directly into your digital audio tracks.

📰 https://news.ycombinator.com/item?id=49061887

🆔 @hackernewsgithubprojects

📦 nikolays/pgsimcity

PGSimCity

PGSimCity is the interactive 3D visualization that finally lets you see exactly how PostgreSQL works under the hood. Instead of dry documentation, it turns database internals into an explorable, living metropolis where every building represents a real mechanism. You can watch client connections stream in, see the shared buffers plaza thrash as pages get evicted, and witness how a single forgotten transaction bloats your storage in real time. It is built for curious developers who want to understand database latency, checkpoints, and replication without drowning in academic jargon. Take a walk through your data directory and master your database today.

📰 https://news.ycombinator.com/item?id=49061947

🆔 @hackernewsgithubprojects

📦 helasaoudi/llm-inspector

llm-inspector

llm-inspector is the terminal utility that finally explains exactly why your GPU memory is full during local artificial intelligence inference. Instead of just showing a flat VRAM number like standard monitoring tools, this clever utility operates like htop specifically for language models. It probes your active processes to break down your GPU memory into clear categories like weights, key-value cache, and workspace. It even projects future savings from quantization strategies like AWQ or FP8 before you modify your setup. You get a clear, measured breakdown of your hardware limits, helping you pinpoint bottlenecks instantly. Try running the inspect command on your active Ollama or vLLM process to finally see where...

📰 https://news.ycombinator.com/item?id=48956776

🆔 @hackernewsgithubprojects

📦 deepender25/edge-drop

Edge-Drop Clipboard

This desktop application turns the edge of your computer screen into a clever, hidden storage shelf for your clipboard. Called edge-drop, it sits completely invisible and click-through on the side of your monitor until you hover your mouse near it, causing a sleek panel to slide out instantly. You can copy images, text, links, or files, and they will collect inside this panel, letting you easily drag them right back out into applications like word processors, design tools, or chat windows. It solves the frustration of constantly switching windows just to copy and paste multiple items, keeping your workflow incredibly fast and natural.

📰 https://news.ycombinator.com/item?id=49078379

🆔 @hackernewsgithubprojects

📦 deerwork-ai/deer-workflow

deer-workflow

You can now define AI agent execution paths entirely in reviewable TypeScript instead of trusting unpredictable LLM conversations. The project deer-workflow is an open-source graph engineering runtime that lets you write clean, deterministic code for your workflow structure while delegating actual semantic tasks to replaceable agent runtimes like Codex and Claude. It comes with a beautiful, phase-aware terminal user interface and even streams a stable JSON event log for automation pipelines. It is a brilliant way to make complex agent behavior observable and structured. If you want reliable, code-first AI workflows, you should definitely check this out.

🆔 @hackernewsgithubprojects

📦 achrefelouafi/oceanthreejs

Oceanthreejs

This open-source water simulator renders a cinematic, endlessly tiling ocean directly in your browser using pure WebGL2 and Three.js with zero build steps or bundlers required. It achieves a near-photoreal look by combining mathematical models: a fast Fourier transform simulation drives the complex wind waves, while multi-layer Gerstner swell functions break up repeating patterns and extend the horizon. The shader pipeline models how real light behaves, calculating physical reflections, depth-based refraction, and even sub-surface scattering that makes wave crests glow from behind. It is the perfect playground for anyone wanting to study cinematic water mathematics in real time.

🆔 @hackernewsgithubprojects

📦 moonshotai/kimi-k3

kimi-k3

Build and run complex engineering projects with minimal human oversight using a massive, open-weight artificial intelligence model. This release presents kimi-k3, a two point eight trillion parameter multimodal model designed to act as an agent for deep research and complex coding tasks. It stands out by handling text, images, and video natively across a massive one-million-token context window, allowing it to navigate large code repositories and coordinate terminal tools easily. By releasing these weights openly, the project makes top-tier frontier intelligence accessible to researchers and developers everywhere. Check out the project to explore this massive leap in open-source AI capabilities.

📰 https://news.ycombinator.com/item?id=49070798

🆔 @hackernewsgithubprojects

📦 moonshotai/moonep

moonep

moonep is the expert parallelism communication library that completely eliminates token load imbalance during mixture of experts training. Standard setups slow down to the pace of the single busiest graphics processor, wasting valuable computing power, but this project solves that by dynamically duplicating hot experts and prefetching them on the fly. By keeping token loads perfectly balanced across all processors, it prevents memory fragmentation and makes sure training runs never crash from running out of memory. It even implements zero-copy communication to write data straight to its final destination, keeping your pipeline incredibly fast. Grab it to keep your hardware running at peak efficiency.

📰 https://news.ycombinator.com/item?id=49071019

🆔 @hackernewsgithubprojects

📦 kevinpbuckley/vibeue

VibeUE

Diagnose and solve performance bottlenecks directly in Unreal Engine five point eight with an intelligent assistant that understands your project. While the engine's built-in AI tools let you build basic assets, they cannot measure how your game actually runs. This project connects directly to Unreal's native model context protocol, giving your AI agents deep diagnostic capabilities to analyze frame rates, run complex traces, and identify whether your game is limited by the graphics card or the processor. It bridges the gap between creative design and technical optimization, letting you fix complex performance issues through natural conversation. Optimize your game's performance with a smarter workflow.

📰 https://news.ycombinator.com/item?id=49079130

🆔 @hackernewsgithubprojects

📦 eldergenix/plato-scientific-research-autonomous-agent

Plato Scientific Research Autonomous Agent

Plato-scientific-research-autonomous-agent is the multi-agent workflow engine that turns raw experimental data into complete, publication-ready research papers. Instead of relying on a single prompt, this clever tool coordinates dedicated agents using explicit state machines to brainstorm ideas, write analysis code, and draft manuscripts. What makes it special is its strict verification loop: it cross-references citations against registries and builds a claim-to-evidence matrix to keep the AI honest. It even packages everything into neat LaTeX documents formatted for specific academic journals. Check it out to see how structured agents can take the tedious formatting and drafting off your plate.

🆔 @hackernewsgithubprojects

📦 wzzheng/sm4rt

SM4RT 4D Reconstruction

Reconstructing dynamic, moving scenes in four dimensions usually requires a messy, tangled web of complex coordinates, but this new tool completely simplifies the math by breaking motion down into clean, structured geometric building blocks. The project is called sm4rt, and it is a fresh approach to four dimensional scene reconstruction that organizes how objects move over time. By decomposing complex motion into neat, simplified latent bases, it bypasses the typical jittery chaos of dynamic tracking. For developers working on spatial computing or robotics, this means you get incredibly clean, stable motion tracking and shape reconstruction directly from video inputs. It is a massive step forward for making machines truly understand...

🆔 @hackernewsgithubprojects

📦 qwen-applications/skill-self-play

Skill Self Play

Skill Self-Play is the training framework that finally helps artificial intelligence models teach themselves advanced logical reasoning and tool-calling skills. Instead of feeding models a fixed, unchanging set of training tasks, this repository lets an AI questioner and an AI solver play against each other to co-evolve. The questioner generates fresh, verified tasks tailored specifically to the solver's current learning limit, pushing the boundaries of what the model can actually handle. It makes training highly targeted, adapting as the model grows smarter. You get vastly stronger reasoning capabilities at training time, while your final model stays simple and fast for everyday use.

🆔 @hackernewsgithubprojects

📦 lyl1015/jarvishub

Jarvis Hub

Jarvis Hub is the canvas-native creative agent harness that finally keeps your entire design workflow visible and interactive in a single workspace. Instead of forcing you into linear chat windows or complex node-based pipelines, this clever system turns an editable canvas into a shared playground for you and AI agents. It maps out text, storyboards, images, and code as visual nodes that you can easily move, link, and edit. Whether you are generating cinematic storyboards, building interactive web pages, or crafting slide presentations, the agent dynamically inspects the canvas to fix errors and build out new versions while preserving all your feedback. Check it out and bring your creative projects...

🆔 @hackernewsgithubprojects

📦 flocknedevastate/augment-code

Augment Code

Augment Code is the AI assistant that actually understands the unique patterns of massive software projects. Instead of just giving generic coding tips, it reads and indexes your entire workspace to understand how your team specifically writes software. This means when you use its chat interface or inline suggestions in your editor, it tailors every line of code to match your existing setup. It plugs directly into tools like Visual Studio Code and JetBrains, saving developers from manually hunting down context across thousands of files. It makes working in giant, complex projects feel as fast and simple as building a small weekend app.

🆔 @hackernewsgithubprojects

📦 wassgha/rescript

Rescript

Rescript is the open source browser app that finally lets you edit video and audio files just by editing text. You simply drop in an audio or video file, and the app transcribes it fully on your device with word-level timestamps. When you delete a word or phrase from the transcript, the app automatically cuts that corresponding section from the media timeline. It runs completely offline using local intelligence, meaning your media never leaves your computer, and you can even strip out filler words like um and uh with a single click. When you are done, you can export your finished cut right from your browser.

📰 https://news.ycombinator.com/item?id=49065779

🆔 @hackernewsgithubprojects

📦 lzy-dot/specflow

SpecFlow

Vision language models can swallow massive amounts of computing power just processing image tokens, but a clever project called specflow uses spectral heat flow to condense those tokens without losing critical visual data. Instead of randomly dropping pixels or running heavy neural network layers to decide what is important, it models token relationships like heat diffusing across a surface. This elegant physics based approach lets you run visual chatbots like Lava much faster on standard hardware because it keeps only the most essential visual information intact. If you want to speed up visual AI models without destroying their accuracy, this implementation is definitely worth checking out.

🆔 @hackernewsgithubprojects

📦 kimberlyii/omni-prune

Omni-Prune

You can now speed up your multimedia artificial intelligence models by over three times without retraining them at all. This is possible because omni-prune intelligently shrinks the massive mountain of data created when models process audio and video together. Usually, processing these synchronized files takes up massive amounts of computer memory and slows down response times. This project solves that by analyzing both your text prompt and the audio peaks to pinpoint exactly which parts of the video are actually relevant. It then discards the unnecessary background data, slashing memory usage while retaining over ninety-nine percent of the model's original accuracy.

🆔 @hackernewsgithubprojects