Media is too big
VIEW IN TELEGRAM
π¦ lins-lab/ultratext_bench
UltraText Bench: The Truth About AI Text
UltraText Bench puts image generators on trial to see if they can actually write real words. It tests if models can render dense English and Chinese text correctly across entire scenes, checking everything from tiny labels to massive posters. The coolest part is that it catches when text looks sharp but says the wrong thing. It uses an automated judge to score accuracy, clarity, and placement, giving you a clear number on whether an AI is truly reliable for text. So the next time an image looks perfect, remember that looking good is not the same as being right, and this tool tells you the difference.
π @hackernewsgithubprojects
UltraText Bench: The Truth About AI Text
UltraText Bench puts image generators on trial to see if they can actually write real words. It tests if models can render dense English and Chinese text correctly across entire scenes, checking everything from tiny labels to massive posters. The coolest part is that it catches when text looks sharp but says the wrong thing. It uses an automated judge to score accuracy, clarity, and placement, giving you a clear number on whether an AI is truly reliable for text. So the next time an image looks perfect, remember that looking good is not the same as being right, and this tool tells you the difference.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ vectorspacelab/arex-2
AREX-2: The AI That Gets Smarter As It Works
Watch an AI agent get genuinely better at a task just by trying again. AREX-2 is a research project that tests if large language models can learn to fix their own mistakes in real-time, round after round, without needing a full retraining. The key finding is that this self-correction skill, built on coding and math problems, actually transfers to deep research tasks like searching the web for answers. This repo gives you the exact tools to run those tests yourself, ensuring every score is tied to a specific model version and data snapshot so you can trust the results.
π @hackernewsgithubprojects
AREX-2: The AI That Gets Smarter As It Works
Watch an AI agent get genuinely better at a task just by trying again. AREX-2 is a research project that tests if large language models can learn to fix their own mistakes in real-time, round after round, without needing a full retraining. The key finding is that this self-correction skill, built on coding and math problems, actually transfers to deep research tasks like searching the web for answers. This repo gives you the exact tools to run those tests yourself, ensuring every score is tied to a specific model version and data snapshot so you can trust the results.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ qlabs-eng/noise
Noise: Pretraining Without Backprop
A team built a speedrun where the only rule is you cannot use backpropagation. They froze a thirty-seven million parameter transformer model and a fixed dataset, then challenged researchers to train it using only forward passes. No gradients, no autograd, just pure estimation. The current best score beats a standard training run, proving that better search algorithms are finally catching up to the default. It is a direct test of whether brute force learning can actually replace the core engine of deep learning. If you care about the future of AI, this is where the experiment is happening. Watch this space, because the records are already breaking.
π @hackernewsgithubprojects
Noise: Pretraining Without Backprop
A team built a speedrun where the only rule is you cannot use backpropagation. They froze a thirty-seven million parameter transformer model and a fixed dataset, then challenged researchers to train it using only forward passes. No gradients, no autograd, just pure estimation. The current best score beats a standard training run, proving that better search algorithms are finally catching up to the default. It is a direct test of whether brute force learning can actually replace the core engine of deep learning. If you care about the future of AI, this is where the experiment is happening. Watch this space, because the records are already breaking.
π @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
π¦ apexdesk/apexdesk
ApexDesk: The Office Suite In One File
ApexDesk is a full office suite that fits inside a single executable file, with zero dependencies. No installation, no cloud account, and no telemetry; it just runs offline on your machine. Written in Rust, it compiles to one tiny native binary that opens instantly, handling documents, spreadsheets, and presentations just like the heavy desktop apps you are used to. Because it is a standalone program, you can drop it on any Windows, Linux, or Android device and start working immediately. It even works in the terminal for quick edits. So if you are done waiting for Office to load, this is the lightweight alternative that actually works.
π @hackernewsgithubprojects
ApexDesk: The Office Suite In One File
ApexDesk is a full office suite that fits inside a single executable file, with zero dependencies. No installation, no cloud account, and no telemetry; it just runs offline on your machine. Written in Rust, it compiles to one tiny native binary that opens instantly, handling documents, spreadsheets, and presentations just like the heavy desktop apps you are used to. Because it is a standalone program, you can drop it on any Windows, Linux, or Android device and start working immediately. It even works in the terminal for quick edits. So if you are done waiting for Office to load, this is the lightweight alternative that actually works.
π @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
π¦ 0xbakeer/tandemllm
Run a 27 Billion Parameter AI Model Locally on One Tiny Box
Turn a single compact desktop into a private AI server that runs massive language models without touching the cloud. Tandem LLM handles the heavy lifting on an N V D G X Spark, serving a twenty-seven billion parameter model with custom math kernels that make it run three times faster than standard setups. The real magic is in how it generates text: a smaller guessing model drafts the next words, and the big model checks them all in one pass, so you get high quality output without waiting for the slow part. It even remembers your long conversations, so restarting the computer doesnβt clear your brain.
π @hackernewsgithubprojects
Run a 27 Billion Parameter AI Model Locally on One Tiny Box
Turn a single compact desktop into a private AI server that runs massive language models without touching the cloud. Tandem LLM handles the heavy lifting on an N V D G X Spark, serving a twenty-seven billion parameter model with custom math kernels that make it run three times faster than standard setups. The real magic is in how it generates text: a smaller guessing model drafts the next words, and the big model checks them all in one pass, so you get high quality output without waiting for the slow part. It even remembers your long conversations, so restarting the computer doesnβt clear your brain.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ tasteskill/tasteskill
Taste Skill: Stop the AI Sludge
Taste Skill forces AI coding agents to build websites that actually look good, fixing the generic, repetitive designs most bots produce by default. It works by handing your AI assistant a specific set of design rules that prioritize strong layout, proper spacing, and readable text over quick, lazy shortcuts. The real magic happens when you let the AI create its own design sketches first, analyze them carefully, and only then write the code to match that vision. This simple shift stops the machine from guessing and makes it act like a real art director.
π @hackernewsgithubprojects
Taste Skill: Stop the AI Sludge
Taste Skill forces AI coding agents to build websites that actually look good, fixing the generic, repetitive designs most bots produce by default. It works by handing your AI assistant a specific set of design rules that prioritize strong layout, proper spacing, and readable text over quick, lazy shortcuts. The real magic happens when you let the AI create its own design sketches first, analyze them carefully, and only then write the code to match that vision. This simple shift stops the machine from guessing and makes it act like a real art director.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ xiayu1987/noobot
Noobot: Your Self-Hosted AI Agent with 3D Animation
Noobot is the self-hosted AI workspace that lets you animate 3D characters just by talking. This open-source tool gives you a private, isolated environment to run AI agents with real-world tools like file management and web browsing. The real surprise here is the character plugin. You can import a 3D model, tell it what to do in plain English, and watch it wave, walk, or jump right in the chat. No coding required. It saves these animations as works you can tweak later. If you want your AI to show, not just say, Noobot is a cool way to see that happen.
π @hackernewsgithubprojects
Noobot: Your Self-Hosted AI Agent with 3D Animation
Noobot is the self-hosted AI workspace that lets you animate 3D characters just by talking. This open-source tool gives you a private, isolated environment to run AI agents with real-world tools like file management and web browsing. The real surprise here is the character plugin. You can import a 3D model, tell it what to do in plain English, and watch it wave, walk, or jump right in the chat. No coding required. It saves these animations as works you can tweak later. If you want your AI to show, not just say, Noobot is a cool way to see that happen.
π @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
π¦ tanakh/plamo-translate-desktop
Translate Anything Offline with PLaMo Translate Desktop
Stop sending your private chats to the cloud just to get a translation. PLaMo Translate Desktop lets you run a powerful AI model right on your own computer, so your words never leave your machine. Once you download the model, you can translate between fifty-three languages completely offline, meaning no Wi-Fi, no subscription, and no data leaks. It handles everything from quick chat messages to entire books, keeping the tone consistent as it goes. You can even plug it into your browser to translate whole websites in place. It is a privacy-first way to communicate across languages without worrying about who is reading over your shoulder.
π @hackernewsgithubprojects
Translate Anything Offline with PLaMo Translate Desktop
Stop sending your private chats to the cloud just to get a translation. PLaMo Translate Desktop lets you run a powerful AI model right on your own computer, so your words never leave your machine. Once you download the model, you can translate between fifty-three languages completely offline, meaning no Wi-Fi, no subscription, and no data leaks. It handles everything from quick chat messages to entire books, keeping the tone consistent as it goes. You can even plug it into your browser to translate whole websites in place. It is a privacy-first way to communicate across languages without worrying about who is reading over your shoulder.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ 34306/doublera1n
Jailbreak Without the Boot Loop
Keep your iPhone running its stock operating system while secretly hosting a second, fully jailbroken environment on the same hardware. This project lets you boot a separate data volume using a small piece of hardware that pwns the device's recovery mode. It solves the annoying problem where traditional jailbreaks force you to reboot constantly or risk bricking your phone into recovery mode. You get a real jailbreak environment with apps and tweaks, but if you just reboot normally, you go straight back to your normal, secure phone.
π @hackernewsgithubprojects
Jailbreak Without the Boot Loop
Keep your iPhone running its stock operating system while secretly hosting a second, fully jailbroken environment on the same hardware. This project lets you boot a separate data volume using a small piece of hardware that pwns the device's recovery mode. It solves the annoying problem where traditional jailbreaks force you to reboot constantly or risk bricking your phone into recovery mode. You get a real jailbreak environment with apps and tweaks, but if you just reboot normally, you go straight back to your normal, secure phone.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ tooonychen/finallyshell
Finallyshell: The SSH Client Without Java
Finallyshell is the SSH client that finally brings the classic FinalShell layout to modern systems without the Java headache. It is a native application written in Go, so you skip the heavy runtime setup and get a lightweight tool that runs smoothly on Mac, Windows, and Linux. The real magic is the interface, which mirrors the familiar vertical session tabs and server monitor layout that power users love, but feels much snappier. It even uses the high-performance Ghostty terminal engine for crisp text rendering. You get all the organization of your favorite tools without the bloat.
π @hackernewsgithubprojects
Finallyshell: The SSH Client Without Java
Finallyshell is the SSH client that finally brings the classic FinalShell layout to modern systems without the Java headache. It is a native application written in Go, so you skip the heavy runtime setup and get a lightweight tool that runs smoothly on Mac, Windows, and Linux. The real magic is the interface, which mirrors the familiar vertical session tabs and server monitor layout that power users love, but feels much snappier. It even uses the high-performance Ghostty terminal engine for crisp text rendering. You get all the organization of your favorite tools without the bloat.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ jandedobbeleer/waku
Waku: Your Custom Windows Desktop
Waku turns Windows into a fully customizable canvas by letting you design your own taskbar, dock, and panels from scratch. Written in a declarative language for Qt Six, it hands developers direct access to system-level features like virtual workspaces, system tray icons, and media controls. This solves the problem of Windows feeling rigid and unchangeable, allowing users to build a shell that fits their exact workflow rather than accepting the default look. It is a developer tool for people who want total control over how their computer interface behaves. If you are tired of the standard Windows bar, Waku gives you the building blocks to create your own desktop experience.
π @hackernewsgithubprojects
Waku: Your Custom Windows Desktop
Waku turns Windows into a fully customizable canvas by letting you design your own taskbar, dock, and panels from scratch. Written in a declarative language for Qt Six, it hands developers direct access to system-level features like virtual workspaces, system tray icons, and media controls. This solves the problem of Windows feeling rigid and unchangeable, allowing users to build a shell that fits their exact workflow rather than accepting the default look. It is a developer tool for people who want total control over how their computer interface behaves. If you are tired of the standard Windows bar, Waku gives you the building blocks to create your own desktop experience.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ robot-agent-exploration/rax
How RAX Teaches Robots by Writing Code
A standard coding agent can teach a physical robot a new task and then compile that lesson into a program that runs two hundred times faster than the agent itself. This project, called Rax, uses an artificial intelligence to drive a robot through trial and error. As the robot explores its environment, it records what works and turns those successful actions into clean, executable code. Once that code is compiled, the slow AI steps back and the robot performs the task on its own at high speed. This approach means the robot improves without needing human demonstrations, and it handles everything from simple arms to complex humanoids.
π @hackernewsgithubprojects
How RAX Teaches Robots by Writing Code
A standard coding agent can teach a physical robot a new task and then compile that lesson into a program that runs two hundred times faster than the agent itself. This project, called Rax, uses an artificial intelligence to drive a robot through trial and error. As the robot explores its environment, it records what works and turns those successful actions into clean, executable code. Once that code is compiled, the slow AI steps back and the robot performs the task on its own at high speed. This approach means the robot improves without needing human demonstrations, and it handles everything from simple arms to complex humanoids.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ reditorsupport/jgd
jgd: The R Plotting Trick
Render your R plots in any web browser without installing a heavy server. JGD, or JSON Graphics Device, is a tiny C library that hooks into Rβs plotting engine to capture drawing instructions as simple JSON. Instead of relying on fragile embedded renderers, it streams those instructions to any compatible client, like the VS Code extension or a standalone Deno server. This lightweight approach makes your graphics incredibly fast, scalable, and easy to export as crisp vectors. Itβs a clever, minimal fix for a problem that used to require a lot of plumbing. If you plot in R, this is the upgrade you didnβt know you needed.
π @hackernewsgithubprojects
jgd: The R Plotting Trick
Render your R plots in any web browser without installing a heavy server. JGD, or JSON Graphics Device, is a tiny C library that hooks into Rβs plotting engine to capture drawing instructions as simple JSON. Instead of relying on fragile embedded renderers, it streams those instructions to any compatible client, like the VS Code extension or a standalone Deno server. This lightweight approach makes your graphics incredibly fast, scalable, and easy to export as crisp vectors. Itβs a clever, minimal fix for a problem that used to require a lot of plumbing. If you plot in R, this is the upgrade you didnβt know you needed.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ techjarves/uncensored-local-ai-multiplatform
Run AI on your phone: Uncensored Local AI Multi-Platform
Run powerful AI models entirely on your phone. Uncensored Local AI Multi-Platform lets you chat with intelligent assistants that never leave your device. It runs open-source models locally on Android, iOS, and desktops, meaning no cloud servers, no monthly fees, and total privacy. The standout feature is that it works completely offline without safety filters. Your data stays private because nothing is sent to a server. It is a great choice for anyone wanting to use AI without relying on the internet or sharing personal information. This app proves you can have high-quality AI right in your pocket.
π @hackernewsgithubprojects
Run AI on your phone: Uncensored Local AI Multi-Platform
Run powerful AI models entirely on your phone. Uncensored Local AI Multi-Platform lets you chat with intelligent assistants that never leave your device. It runs open-source models locally on Android, iOS, and desktops, meaning no cloud servers, no monthly fees, and total privacy. The standout feature is that it works completely offline without safety filters. Your data stays private because nothing is sent to a server. It is a great choice for anyone wanting to use AI without relying on the internet or sharing personal information. This app proves you can have high-quality AI right in your pocket.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ xoxor4d/gta4-rtx
GTA 4 RTX Remix Mod
Turn a 2008 classic into a modern visual showcase. This community patch hooks into NVIDIAβs RTX Remix tech to bring real time ray tracing to Grand Theft Auto IV. It lets you see light bounce off streets, wet surfaces reflect the sky, and volumetric clouds drift overhead in ways the original game never supported. The developer focused heavily on fixing performance bottlenecks so the extra detail does not crash your frame rate. It is not a full remaster, just a bridge for the engine to handle modern lighting effects.
π @hackernewsgithubprojects
GTA 4 RTX Remix Mod
Turn a 2008 classic into a modern visual showcase. This community patch hooks into NVIDIAβs RTX Remix tech to bring real time ray tracing to Grand Theft Auto IV. It lets you see light bounce off streets, wet surfaces reflect the sky, and volumetric clouds drift overhead in ways the original game never supported. The developer focused heavily on fixing performance bottlenecks so the extra detail does not crash your frame rate. It is not a full remaster, just a bridge for the engine to handle modern lighting effects.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ huamidev/gestaltx
GestaltX: Unlock Hidden iPhone Features
GestaltX is the iOS tool that finally lets you edit your phone's hidden configuration file directly on the device. It solves the problem of locked features by giving you precise control over the internal system settings that Apple usually hides from regular users. The most surprising part is how safely it handles these changes. Before you touch anything, it automatically saves a backup of your original settings. This means you can experiment with things like region codes or display modes without the panic of breaking your phone. If a change does not work, you can simply tap a button to restore your device to its exact previous state.
π @hackernewsgithubprojects
GestaltX: Unlock Hidden iPhone Features
GestaltX is the iOS tool that finally lets you edit your phone's hidden configuration file directly on the device. It solves the problem of locked features by giving you precise control over the internal system settings that Apple usually hides from regular users. The most surprising part is how safely it handles these changes. Before you touch anything, it automatically saves a backup of your original settings. This means you can experiment with things like region codes or display modes without the panic of breaking your phone. If a change does not work, you can simply tap a button to restore your device to its exact previous state.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ php-debugger/php-debugger
PHP Debugger: The Zero-Overhead Debugging Tool
PHP Debugger lets you keep your code debugger switched on permanently without slowing down your application. Built on the foundation of Xdebug but stripped down to the essentials, it removes the heavy profiling and tracing features that usually cause performance hits. This means you can step through your code, inspect variables, and fix bugs with near-zero overhead when you are not actively debugging. It even supports AI agents, allowing tools like Claude to set breakpoints and debug your code automatically, just like you would. Since it speaks the standard debugging protocol, it works seamlessly with your existing editor setups.
π @hackernewsgithubprojects
PHP Debugger: The Zero-Overhead Debugging Tool
PHP Debugger lets you keep your code debugger switched on permanently without slowing down your application. Built on the foundation of Xdebug but stripped down to the essentials, it removes the heavy profiling and tracing features that usually cause performance hits. This means you can step through your code, inspect variables, and fix bugs with near-zero overhead when you are not actively debugging. It even supports AI agents, allowing tools like Claude to set breakpoints and debug your code automatically, just like you would. Since it speaks the standard debugging protocol, it works seamlessly with your existing editor setups.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ arcanima/comfyui-anima-ref-adapter
Anima Reference Adapter for ComfyUI
Drop a single photo into ComfyUI and watch the Anima Reference Adapter lock in that exact character's face, hair, and outfit, even as you type in a completely different scene. The surprising part is that it does not copy the pose or background, only the identity. You describe the action and location in your prompt, while the model pulls the visual style and specific features directly from your reference image. It is so simple that you only add one node to your existing workflow, and you can even stitch two photos together to place two distinct characters in one picture.
π @hackernewsgithubprojects
Anima Reference Adapter for ComfyUI
Drop a single photo into ComfyUI and watch the Anima Reference Adapter lock in that exact character's face, hair, and outfit, even as you type in a completely different scene. The surprising part is that it does not copy the pose or background, only the identity. You describe the action and location in your prompt, while the model pulls the visual style and specific features directly from your reference image. It is so simple that you only add one node to your existing workflow, and you can even stitch two photos together to place two distinct characters in one picture.
π @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
π¦ samir-sarwar/samir-vllm
How Samir-VLLM Builds LLM Inference From Scratch
Samir-VLLM is the educational project that builds a working large language model inference engine entirely from scratch. Instead of relying on heavy pre-built libraries, this C++ and CUDA project lets you see exactly how a text prompt transforms into generated words by running every step directly on the graphics processor. The standout feature is the comprehensive collection of over six hours of video tutorials paired with eighty-five pages of handwritten notes. This resource guides you through the entire journey, from loading model weights to executing complex attention mechanisms, making the invisible math of artificial intelligence completely visible.
π° https://news.ycombinator.com/item?id=50424199
π @hackernewsgithubprojects
How Samir-VLLM Builds LLM Inference From Scratch
Samir-VLLM is the educational project that builds a working large language model inference engine entirely from scratch. Instead of relying on heavy pre-built libraries, this C++ and CUDA project lets you see exactly how a text prompt transforms into generated words by running every step directly on the graphics processor. The standout feature is the comprehensive collection of over six hours of video tutorials paired with eighty-five pages of handwritten notes. This resource guides you through the entire journey, from loading model weights to executing complex attention mechanisms, making the invisible math of artificial intelligence completely visible.
π° https://news.ycombinator.com/item?id=50424199
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ nyabsi/openvr-spaceoverride
OpenVR Space Override
OpenVR Space Override is the virtual reality driver that kills headset drift by borrowing precision from external base stations. It fixes a common problem where cheap headsets slowly slide out of place. Instead of just guessing your position, it mounts a tracker to your headset and uses that stable signal to lock your view in place. This means your head and controllers stay perfectly aligned for hours, without that annoying sliding effect. It is a clever workaround for anyone mixing different tracking systems who wants to stop fighting the software.
π @hackernewsgithubprojects
OpenVR Space Override
OpenVR Space Override is the virtual reality driver that kills headset drift by borrowing precision from external base stations. It fixes a common problem where cheap headsets slowly slide out of place. Instead of just guessing your position, it mounts a tracker to your headset and uses that stable signal to lock your view in place. This means your head and controllers stay perfectly aligned for hours, without that annoying sliding effect. It is a clever workaround for anyone mixing different tracking systems who wants to stop fighting the software.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ ifreemz/msfs2024-glass-cockpit-dlss-fix
The Glass Cockpit DLSS Fix for MSFS 2024
A small add-on for Microsoft Flight Simulator 2024 permanently erases the blurry ghosting that appears on digital cockpit screens when you use upscalers like DLSS or FSR. Usually, these AI tools smear speed and altitude numbers because they are designed for video, not precise instruments. This tool intercepts the graphics driver to redraw those specific screens with perfect clarity, ensuring your readouts stay sharp whether you are flying on a monitor or wearing a virtual reality headset. It works as a simple plug-in that fixes the visual noise without slowing down your game.
π @hackernewsgithubprojects
The Glass Cockpit DLSS Fix for MSFS 2024
A small add-on for Microsoft Flight Simulator 2024 permanently erases the blurry ghosting that appears on digital cockpit screens when you use upscalers like DLSS or FSR. Usually, these AI tools smear speed and altitude numbers because they are designed for video, not precise instruments. This tool intercepts the graphics driver to redraw those specific screens with perfect clarity, ensuring your readouts stay sharp whether you are flying on a monitor or wearing a virtual reality headset. It works as a simple plug-in that fixes the visual noise without slowing down your game.
π @hackernewsgithubprojects