This media is not supported in your browser
VIEW IN TELEGRAM
π¦ amazon-far/ssio
SSIO: Training Robots Without GPS
SSIO allows robots to learn their own position from raw sensor data alone, with absolutely zero GPS or motion tracking labels. Typically, teaching a machine to track its movement requires expensive, manually drawn paths, but this project changes that. It feeds a neural network simple motion readings from a device called an IMU, letting the system figure out velocity and position through a clever self-supervised trick. It works on everything from tiny drones to walking humanoids. The result is a lightweight navigation tool that runs on standard hardware without needing a single external map.
π @hackernewsgithubprojects
SSIO: Training Robots Without GPS
SSIO allows robots to learn their own position from raw sensor data alone, with absolutely zero GPS or motion tracking labels. Typically, teaching a machine to track its movement requires expensive, manually drawn paths, but this project changes that. It feeds a neural network simple motion readings from a device called an IMU, letting the system figure out velocity and position through a clever self-supervised trick. It works on everything from tiny drones to walking humanoids. The result is a lightweight navigation tool that runs on standard hardware without needing a single external map.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ kaiserwholearns/reasoningeffortllm
The 42x Cost of Overthinking
Turning up a model's reasoning dial to its maximum doesn't just make it smarter; it makes it dramatically more expensive. A new study from Johns Hopkins and Together AI tracked eight different artificial intelligence models across four tough benchmarks to see exactly what happens when you push that effort setting all the way up. The results are startling: on tasks the model solves at every level, the highest setting costs up to forty-two times more than the lowest. Even wilder, cranking up the effort doesn't guarantee better results. Often, a model running on high effort completely fails a task that the same model easily solved on low effort.
π @hackernewsgithubprojects
The 42x Cost of Overthinking
Turning up a model's reasoning dial to its maximum doesn't just make it smarter; it makes it dramatically more expensive. A new study from Johns Hopkins and Together AI tracked eight different artificial intelligence models across four tough benchmarks to see exactly what happens when you push that effort setting all the way up. The results are startling: on tasks the model solves at every level, the highest setting costs up to forty-two times more than the lowest. Even wilder, cranking up the effort doesn't guarantee better results. Often, a model running on high effort completely fails a task that the same model easily solved on low effort.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ maxdevv/lean-qwenflash
Lean-QwenFlash: 262k Context on a Laptop
A hundred and twenty-five billion parameter language model usually demands forty-eight gigabytes of system memory just to stay awake, but Lean-QwenFlash runs it on a single high-end graphics card with only eight gigabytes of regular computer memory. It pulls expert weights directly from your solid state drive into the graphics card as it needs them, skipping the slow main memory entirely. This lets a consumer gaming rig process two hundred and sixty-two thousand tokens of context at speeds that rival much larger workstations. You get near-instant response times for long documents without buying a server rack.
π @hackernewsgithubprojects
Lean-QwenFlash: 262k Context on a Laptop
A hundred and twenty-five billion parameter language model usually demands forty-eight gigabytes of system memory just to stay awake, but Lean-QwenFlash runs it on a single high-end graphics card with only eight gigabytes of regular computer memory. It pulls expert weights directly from your solid state drive into the graphics card as it needs them, skipping the slow main memory entirely. This lets a consumer gaming rig process two hundred and sixty-two thousand tokens of context at speeds that rival much larger workstations. You get near-instant response times for long documents without buying a server rack.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ potionui/potionui
PotionUI: Your Private AI Studio
Turn your own hardware into a complete creative studio. PotionUI is a self-hosted platform that lets you generate images, videos, and audio using AI models right on your own machine. The real magic here is the ability to control everything locally. You get a full interface to manage models, tweak settings, and run generations without sending your data to any cloud service. It supports a wide range of backends, including local GPUs and remote workers, giving you total freedom. This is the tool for anyone who wants the power of AI without the subscription fees or the privacy risks. You own the output, and you own the process.
π @hackernewsgithubprojects
PotionUI: Your Private AI Studio
Turn your own hardware into a complete creative studio. PotionUI is a self-hosted platform that lets you generate images, videos, and audio using AI models right on your own machine. The real magic here is the ability to control everything locally. You get a full interface to manage models, tweak settings, and run generations without sending your data to any cloud service. It supports a wide range of backends, including local GPUs and remote workers, giving you total freedom. This is the tool for anyone who wants the power of AI without the subscription fees or the privacy risks. You own the output, and you own the process.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ mrdesjardins/audiorouter
AudioRouter: Draw Your Sound
Your computer can mix audio from ten different apps at once without you touching a single slider, all laid out on a visual canvas you can literally read. AudioRouter is a free, open-source tool for Windows that treats sound like a wiring diagram instead of a tangled mess of settings. You drag lines from your microphone to a noise gate, an equalizer, and a compressor, then send that clean voice to Discord, your headphones, and a backup recorder all at the same time. It even lets you stream game audio from one PC to another over your home network without extra software.
π @hackernewsgithubprojects
AudioRouter: Draw Your Sound
Your computer can mix audio from ten different apps at once without you touching a single slider, all laid out on a visual canvas you can literally read. AudioRouter is a free, open-source tool for Windows that treats sound like a wiring diagram instead of a tangled mess of settings. You drag lines from your microphone to a noise gate, an equalizer, and a compressor, then send that clean voice to Discord, your headphones, and a backup recorder all at the same time. It even lets you stream game audio from one PC to another over your home network without extra software.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ tanujnotes/supernova
Supernova: The Offline Launcher
Supernova is an open source phone launcher that works entirely offline, meaning your entire home screen functions without an internet connection. It replaces the standard app interface with a system that finds apps, contacts, and files instantly, and even calculates math right in the search bar. Since it does not rely on cloud services, your data stays on your device, giving you full control over speed and privacy. This project offers a fast, private alternative to launchers that often track your behavior. If you want a responsive home screen that respects your connection limits, Supernova gives you that power.
π @hackernewsgithubprojects
Supernova: The Offline Launcher
Supernova is an open source phone launcher that works entirely offline, meaning your entire home screen functions without an internet connection. It replaces the standard app interface with a system that finds apps, contacts, and files instantly, and even calculates math right in the search bar. Since it does not rely on cloud services, your data stays on your device, giving you full control over speed and privacy. This project offers a fast, private alternative to launchers that often track your behavior. If you want a responsive home screen that respects your connection limits, Supernova gives you that power.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ sam1983aing/flat-to-3d
Flat to 3D
Flat to 3D is the collection that turns flat architectural drawings into real buildings you can walk through. The most surprising thing it does is show that a flat 2D drawing can literally become the 3D model itself. In five of the seven demos, the flat image is actually the real house viewed through a very long lens, so it looks flat until the camera opens up and reveals the full building. Each demo uses plain JavaScript to build every wall and piece of furniture in code, with no image files or 3D models.
π @hackernewsgithubprojects
Flat to 3D
Flat to 3D is the collection that turns flat architectural drawings into real buildings you can walk through. The most surprising thing it does is show that a flat 2D drawing can literally become the 3D model itself. In five of the seven demos, the flat image is actually the real house viewed through a very long lens, so it looks flat until the camera opens up and reveals the full building. Each demo uses plain JavaScript to build every wall and piece of furniture in code, with no image files or 3D models.
π @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
π¦ miaai-lab/deepseek-v4.1-flash-1017-4x-dgx-sparks-tensorfold
DeepSeek V4.1 Flash on 4 DGX Sparks
DeepSeek V4.1 Flash on four DGX Sparks is the local AI setup that serves a billion-parameter model without a network switch. It takes four standalone GPU boxes, links them in a simple cable ring, and splits the massive model across all of them so they act as one giant machine. The cool part is that it handles over a million words of context simultaneously, meaning you can feed it entire codebases or books and it remembers everything perfectly. You get an API that accepts text, images, and video, plus a built-in speed booster that guesses the next words to make responses faster.
π @hackernewsgithubprojects
DeepSeek V4.1 Flash on 4 DGX Sparks
DeepSeek V4.1 Flash on four DGX Sparks is the local AI setup that serves a billion-parameter model without a network switch. It takes four standalone GPU boxes, links them in a simple cable ring, and splits the massive model across all of them so they act as one giant machine. The cool part is that it handles over a million words of context simultaneously, meaning you can feed it entire codebases or books and it remembers everything perfectly. You get an API that accepts text, images, and video, plus a built-in speed booster that guesses the next words to make responses faster.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ sahenjp/rustxng
RustXNG: The Rust Powerhouse
RustXNG is the private metasearch engine that finally rivals Python speed. It takes the beloved SearXNG tool and rewrites the entire core in Rust, delivering lightning-fast query times without sacrificing the privacy you expect. Instead of relying on heavy web scrapers, it queries over one hundred distinct data sources simultaneously, including academic archives and code repositories. This means you get precise, relevant results in milliseconds while keeping your search history strictly on your own hardware. If you value both speed and sovereignty, this project is the one to watch.
π @hackernewsgithubprojects
RustXNG: The Rust Powerhouse
RustXNG is the private metasearch engine that finally rivals Python speed. It takes the beloved SearXNG tool and rewrites the entire core in Rust, delivering lightning-fast query times without sacrificing the privacy you expect. Instead of relying on heavy web scrapers, it queries over one hundred distinct data sources simultaneously, including academic archives and code repositories. This means you get precise, relevant results in milliseconds while keeping your search history strictly on your own hardware. If you value both speed and sovereignty, this project is the one to watch.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ astral-group/benchmark-difficulty-measure
Benchmark Difficulty Measure: Filling the Gaps
Benchmark Difficulty Measure treats every AI test as a map of problem difficulty, exposing blind spots where models can improve but never get scored. It calculates two key metrics to reveal if a benchmark is missing easy or hard questions, then uses a tool called BenchFill to generate brand-new problems specifically for those empty zones. By plugging these holes with new items, the system ensures that if a model actually gets smarter in a specific area, the test will finally show it, making comparisons between different AI models much fairer and more accurate.
π @hackernewsgithubprojects
Benchmark Difficulty Measure: Filling the Gaps
Benchmark Difficulty Measure treats every AI test as a map of problem difficulty, exposing blind spots where models can improve but never get scored. It calculates two key metrics to reveal if a benchmark is missing easy or hard questions, then uses a tool called BenchFill to generate brand-new problems specifically for those empty zones. By plugging these holes with new items, the system ensures that if a model actually gets smarter in a specific area, the test will finally show it, making comparisons between different AI models much fairer and more accurate.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ improsocial/impro
Impro: The Bluesky App That Lets You Build Your Own
Stop settling for the default feed. Impro lets you hack your own social media experience on the web, right in your browser. Think of it as the Obsidian of Bluesky, a sleek client where you install community plugins to add custom features. You are no longer stuck with whatever the app decides to show you. Instead, you grab a tool that lets you inject new menus, change how posts look, or even build your own unique feeds. It is built from scratch, so it feels fast and feels like it was made by people who actually use the app. No heavy frameworks, just clean code that respects your time.
π @hackernewsgithubprojects
Impro: The Bluesky App That Lets You Build Your Own
Stop settling for the default feed. Impro lets you hack your own social media experience on the web, right in your browser. Think of it as the Obsidian of Bluesky, a sleek client where you install community plugins to add custom features. You are no longer stuck with whatever the app decides to show you. Instead, you grab a tool that lets you inject new menus, change how posts look, or even build your own unique feeds. It is built from scratch, so it feels fast and feels like it was made by people who actually use the app. No heavy frameworks, just clean code that respects your time.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ jandej/e486-web
Run DOOM in Your Browser with This Retro PC
Run classic DOS games like DOOM directly in your browser, without any installation or virtual machine. E486-web builds a complete, interactive computer from the 1990s using only JavaScript. You see the physical motherboard, can plug in different expansion cards, and watch the boot process on a simulated CRT monitor. It boots real operating systems like FreeDOS and runs unmodified software from that era. This project is interesting because it lets you experience the hardware logic of a vintage PC, seeing exactly how the system interacts with its own components. It turns a static webpage into a living, breathing retro computer.
π @hackernewsgithubprojects
Run DOOM in Your Browser with This Retro PC
Run classic DOS games like DOOM directly in your browser, without any installation or virtual machine. E486-web builds a complete, interactive computer from the 1990s using only JavaScript. You see the physical motherboard, can plug in different expansion cards, and watch the boot process on a simulated CRT monitor. It boots real operating systems like FreeDOS and runs unmodified software from that era. This project is interesting because it lets you experience the hardware logic of a vintage PC, seeing exactly how the system interacts with its own components. It turns a static webpage into a living, breathing retro computer.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ scribe-ide/scribe
Scribe: The Local LaTeX Editor for Scientists
Keep your research papers, Python notebooks, and generated figures in one local folder with Scribe. This desktop editor solves the messy handoff between analysis and writing. When you run your notebook, Scribe automatically injects the resulting plots and exact numbers directly into your LaTeX document. Everything compiles right on your machine, so your data never leaves your computer. You get a complete, private workspace where your code and your paper always stay in sync. If you write scientific papers, this is the tool that keeps your workflow honest and entirely under your control. Give it a try.
π @hackernewsgithubprojects
Scribe: The Local LaTeX Editor for Scientists
Keep your research papers, Python notebooks, and generated figures in one local folder with Scribe. This desktop editor solves the messy handoff between analysis and writing. When you run your notebook, Scribe automatically injects the resulting plots and exact numbers directly into your LaTeX document. Everything compiles right on your machine, so your data never leaves your computer. You get a complete, private workspace where your code and your paper always stay in sync. If you write scientific papers, this is the tool that keeps your workflow honest and entirely under your control. Give it a try.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ wmahfoudh/okf
okf
okf is the practical guide that turns abstract knowledge frameworks into working reality. It stops being a theory and starts being a system you can actually build. The project maps out every step from initial concept to final deployment, handling the messy lifecycle challenges that usually break these projects. It gives you a clear, tested path so you donβt have to guess how to keep your knowledge structure alive and useful over time. Instead of wondering if your setup will hold up, you get a proven structure that handles the real-world wear and tear of ongoing maintenance.
π @hackernewsgithubprojects
okf
okf is the practical guide that turns abstract knowledge frameworks into working reality. It stops being a theory and starts being a system you can actually build. The project maps out every step from initial concept to final deployment, handling the messy lifecycle challenges that usually break these projects. It gives you a clear, tested path so you donβt have to guess how to keep your knowledge structure alive and useful over time. Instead of wondering if your setup will hold up, you get a proven structure that handles the real-world wear and tear of ongoing maintenance.
π @hackernewsgithubprojects
Media is too big
VIEW IN TELEGRAM
π¦ swedishembedded/brain
brain: AI for your pocket
You can run massive AI models right on your own phone, not just in a data center. The open-source platform called brain brings high-end training and inference to everyday hardware by writing the heavy lifting directly in the programming language Rust. This solves a massive problem: it lets you run sophisticated tools like language models and image generators locally on devices like a Raspberry Pi or a standard laptop. You get the power of big artificial intelligence without waiting for a cloud connection or paying monthly fees. It is a complete toolkit for developers who want to build smart apps that actually work on the device in their hand.
π @hackernewsgithubprojects
brain: AI for your pocket
You can run massive AI models right on your own phone, not just in a data center. The open-source platform called brain brings high-end training and inference to everyday hardware by writing the heavy lifting directly in the programming language Rust. This solves a massive problem: it lets you run sophisticated tools like language models and image generators locally on devices like a Raspberry Pi or a standard laptop. You get the power of big artificial intelligence without waiting for a cloud connection or paying monthly fees. It is a complete toolkit for developers who want to build smart apps that actually work on the device in their hand.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ alexw5702-afk/krea2-anypaint
Krea 2 AnyPaint
You can now paint directly onto the edges of a photo, extending the canvas in any direction without losing the original quality. Krea 2 AnyPaint brings this arbitrary-mask editing to ComfyUI, letting you add or remove content wherever you want, not just inside the frame. It preserves your existing pixels perfectly while generating new areas that match the style and lighting. This native integration means you get precise control over mixed editing tasks like changing a background or completing a cropped subject. If you work with ComfyUI, this is the missing piece for true flexible image manipulation.
π @hackernewsgithubprojects
Krea 2 AnyPaint
You can now paint directly onto the edges of a photo, extending the canvas in any direction without losing the original quality. Krea 2 AnyPaint brings this arbitrary-mask editing to ComfyUI, letting you add or remove content wherever you want, not just inside the frame. It preserves your existing pixels perfectly while generating new areas that match the style and lighting. This native integration means you get precise control over mixed editing tasks like changing a background or completing a cropped subject. If you work with ComfyUI, this is the missing piece for true flexible image manipulation.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ cvlab-kaist/l3g-world
L3G-World
A team at K A I S T built a system that generates infinite, real-time worlds without losing track of what it has already seen. By using a persistent memory made of tiny 3D shapes, it remembers exactly where objects are, so the scene stays consistent no matter how far you explore. This solves the common problem where generated environments flicker or change unexpectedly when you look back. It is useful because it allows for stable, long-duration virtual spaces that feel physically real. The key takeaway is that smart memory, not just powerful generation, is what makes these digital worlds feel truly infinite.
π @hackernewsgithubprojects
L3G-World
A team at K A I S T built a system that generates infinite, real-time worlds without losing track of what it has already seen. By using a persistent memory made of tiny 3D shapes, it remembers exactly where objects are, so the scene stays consistent no matter how far you explore. This solves the common problem where generated environments flicker or change unexpectedly when you look back. It is useful because it allows for stable, long-duration virtual spaces that feel physically real. The key takeaway is that smart memory, not just powerful generation, is what makes these digital worlds feel truly infinite.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ cobusgreyling/grok-bot-templates
Grok Bot Templates: 49 Ready-to-Use AI Agents
Forty-nine pre-written AI agent profiles, each with a score that proves they are ready for work, are sitting in one place waiting for you to copy them. This project solves the problem of blank pages by giving you a catalog of specific jobs, like bug tracking or inbox triage, where you simply paste the profile into your account and start working. It includes clear rules on what the bot should never do, ensuring it stays safe and helpful without you needing to write complex code. It is essentially a library of smart, reliable helpers for your digital workflow.
π @hackernewsgithubprojects
Grok Bot Templates: 49 Ready-to-Use AI Agents
Forty-nine pre-written AI agent profiles, each with a score that proves they are ready for work, are sitting in one place waiting for you to copy them. This project solves the problem of blank pages by giving you a catalog of specific jobs, like bug tracking or inbox triage, where you simply paste the profile into your account and start working. It includes clear rules on what the bot should never do, ensuring it stays safe and helpful without you needing to write complex code. It is essentially a library of smart, reliable helpers for your digital workflow.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ quantumorb/quantorb
QuantumOrb: The Memecoin Truth Serum
QuantumOrb is the memecoin scanner that finally separates real holders from fake whales. You know that annoying feeling when you see twenty wallets holding a token, but you just know those are all the same guy? That is exactly what this tool fixes. It digs into the blockchain to find hidden links between addresses, merging them into one big player so you can see the truth. If that one big holder sold everything today, the price would crash by sixty-seven percent. It also checks if the creator can still pull the rug. It gives you a simple score from zero to one hundred.
π @hackernewsgithubprojects
QuantumOrb: The Memecoin Truth Serum
QuantumOrb is the memecoin scanner that finally separates real holders from fake whales. You know that annoying feeling when you see twenty wallets holding a token, but you just know those are all the same guy? That is exactly what this tool fixes. It digs into the blockchain to find hidden links between addresses, merging them into one big player so you can see the truth. If that one big holder sold everything today, the price would crash by sixty-seven percent. It also checks if the creator can still pull the rug. It gives you a simple score from zero to one hundred.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ yihaohu0118/chronoreplay
ChronoReplay: The AI That Thinks in Time
ChronoReplay is the spatial intelligence model that finally understands the order of events. Instead of forcing a vision system to guess everything at once, this project keeps time in chronological order so the computer sees exactly what happened before making a move. It handles different types of questions together without losing track of the sequence. This approach lets the model reason through space and time in a way that matches how we actually see the world. You get a smarter AI that truly understands what just happened, not just what is in front of it right now.
π @hackernewsgithubprojects
ChronoReplay: The AI That Thinks in Time
ChronoReplay is the spatial intelligence model that finally understands the order of events. Instead of forcing a vision system to guess everything at once, this project keeps time in chronological order so the computer sees exactly what happened before making a move. It handles different types of questions together without losing track of the sequence. This approach lets the model reason through space and time in a way that matches how we actually see the world. You get a smarter AI that truly understands what just happened, not just what is in front of it right now.
π @hackernewsgithubprojects