This media is not supported in your browser
VIEW IN TELEGRAM
π¦ tsukushi-gate/griffith
Griffith: Neural Nets Built From Logic Gates
Griffith lets you train a neural network where every neuron is a lookup table of Boolean gates, then collapse it into pure hardware-ready logic. The trick is that during training the gate is a soft mixture so gradients can flow, and at eval time it snaps to a hard Boolean circuit with no loss of accuracy β like a brain that learns, then locks into a circuit. Once trained, Griffith exports that circuit into Verilog or VHDL, so the model can run on an FPGA instead of a GPU. It even ships models for classification, language, detection, and more, all built from gates.
π @hackernewsgithubprojects
Griffith: Neural Nets Built From Logic Gates
Griffith lets you train a neural network where every neuron is a lookup table of Boolean gates, then collapse it into pure hardware-ready logic. The trick is that during training the gate is a soft mixture so gradients can flow, and at eval time it snaps to a hard Boolean circuit with no loss of accuracy β like a brain that learns, then locks into a circuit. Once trained, Griffith exports that circuit into Verilog or VHDL, so the model can run on an FPGA instead of a GPU. It even ships models for classification, language, detection, and more, all built from gates.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ sous-labs/songe
Songe: A Local Model That Sleeps
Your model is asleep, and itβs learning on its own. Songe turns your idle computer into a training engine for your local AI model, using the rhythm of your day. While you work, it watches what happens after the model acts β the tool result, the error, the next message β and when you step away, it turns those lessons into a smarter version of your model. It does this without any cloud, telemetry, or accounts. The coolest part is the 'Dream': instead of just waiting, the model rehearses what went wrong, practices on synthetic scenes, and only swaps in a new version if it clearly passes checks.
π @hackernewsgithubprojects
Songe: A Local Model That Sleeps
Your model is asleep, and itβs learning on its own. Songe turns your idle computer into a training engine for your local AI model, using the rhythm of your day. While you work, it watches what happens after the model acts β the tool result, the error, the next message β and when you step away, it turns those lessons into a smarter version of your model. It does this without any cloud, telemetry, or accounts. The coolest part is the 'Dream': instead of just waiting, the model rehearses what went wrong, practices on synthetic scenes, and only swaps in a new version if it clearly passes checks.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ deathshotxd/sip2shell
SIP2Shell
SIP2Shell is the small security tool that turns one text-message-style SIP request into running commands on a vulnerable FreeSWITCH server, and its sharpest strength is that it does this with a single packet and a real before-and-after lab. FreeSWITCH is the software behind many phone systems, and the bug it exposes is a missing guard on a chat address: when a SIP MESSAGE arrives with an address like api plus a command, the server reads the left side as a protocol name and hands the rest to a shell-backed API, so one UDP packet can run commands and send the answer back.
π @hackernewsgithubprojects
SIP2Shell
SIP2Shell is the small security tool that turns one text-message-style SIP request into running commands on a vulnerable FreeSWITCH server, and its sharpest strength is that it does this with a single packet and a real before-and-after lab. FreeSWITCH is the software behind many phone systems, and the bug it exposes is a missing guard on a chat address: when a SIP MESSAGE arrives with an address like api plus a command, the server reads the left side as a protocol name and hands the rest to a shell-backed API, so one UDP packet can run commands and send the answer back.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ modelcontextprotocol/ext-events
ext-events
The incubation space that finally lets servers push updates instead of make clients wait. ext-events is the Tasks and Events Working Group inside the Model Context Protocol project, and its one defining strength is making servers talk first. Today, clients have to keep asking or hold a connection open just to notice something changed; this group is designing a standard way for a server to send a callback when new data is ready, with ordering rules that should work the same across different connection styles.
π @hackernewsgithubprojects
ext-events
The incubation space that finally lets servers push updates instead of make clients wait. ext-events is the Tasks and Events Working Group inside the Model Context Protocol project, and its one defining strength is making servers talk first. Today, clients have to keep asking or hold a connection open just to notice something changed; this group is designing a standard way for a server to send a callback when new data is ready, with ordering rules that should work the same across different connection styles.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ thibault-gillard/box2boxcam-football-vision
box2boxcam: an iPhone gimbal becomes your football cameraman
Here is a clever project that turns a phone on a small gimbal into a self-driving cameraman for amateur football matches. It is not a fancy camera; it is a small, fine-tuned detector that finds the tiny, hard-to-spot ball, plus a simple rule that keeps players in frame and moves the phone like a human cameraman. The cool part is that it is cheap, open-source, and runs on the phone itself, no subscription or big hardware. The catch is honest: it does not find every ball, especially small ones in the distance, but it does a lot better than a stock model and is fun to watch.
π @hackernewsgithubprojects
box2boxcam: an iPhone gimbal becomes your football cameraman
Here is a clever project that turns a phone on a small gimbal into a self-driving cameraman for amateur football matches. It is not a fancy camera; it is a small, fine-tuned detector that finds the tiny, hard-to-spot ball, plus a simple rule that keeps players in frame and moves the phone like a human cameraman. The cool part is that it is cheap, open-source, and runs on the phone itself, no subscription or big hardware. The catch is honest: it does not find every ball, especially small ones in the distance, but it does a lot better than a stock model and is fun to watch.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ devohmycode/skills-atlas
Skills Atlas
Skills Atlas is the one-stop shop for finding, checking, and installing Agent Skills. Instead of hunting through scattered repos, you get a searchable catalog of 75,000 public skills organized by theme, all in your terminal or browser. The clever part is the pre-install risk check: it flags skills that download and run code, read credential files, or send data to servers before you ever install them. It also detects impostor names that look like popular skills, so you know exactly what a skill does before it touches your machine. You can browse, tick what you want, review the risks, and install together.
π @hackernewsgithubprojects
Skills Atlas
Skills Atlas is the one-stop shop for finding, checking, and installing Agent Skills. Instead of hunting through scattered repos, you get a searchable catalog of 75,000 public skills organized by theme, all in your terminal or browser. The clever part is the pre-install risk check: it flags skills that download and run code, read credential files, or send data to servers before you ever install them. It also detects impostor names that look like popular skills, so you know exactly what a skill does before it touches your machine. You can browse, tick what you want, review the risks, and install together.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ encidxvii/claude-reverse-engineering-skill
Claude Reverse Engineering Skill
Imagine handing an AI agent a mystery binary and watching it figure out what the file really is, without ever running it. The Claude Reverse Engineering Skill gives Claude Code and other AI coding agents a professional reverse-engineering workflow that identifies files from their raw bytes, checks which tools are installed, decompiles code headlessly, and maps the network surface. It covers native executables, Android and iOS apps, .NET and Java, Electron and WebAssembly, firmware, and malware triage, all safely isolated with no network access.
π @hackernewsgithubprojects
Claude Reverse Engineering Skill
Imagine handing an AI agent a mystery binary and watching it figure out what the file really is, without ever running it. The Claude Reverse Engineering Skill gives Claude Code and other AI coding agents a professional reverse-engineering workflow that identifies files from their raw bytes, checks which tools are installed, decompiles code headlessly, and maps the network surface. It covers native executables, Android and iOS apps, .NET and Java, Electron and WebAssembly, firmware, and malware triage, all safely isolated with no network access.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ laravel/ai
Laravel AI
Let your Laravel app talk to every AI provider without juggling ten different SDKs. This SDK gives you one simple, Laravel-friendly way to build smart agents that can use tools, remember conversations, and return clean structured data β plus generate images, audio, and embeddings. The coolest part: you can give an agent a job, let it think and act step by step, and even pause to ask you for approval before it does something risky. Then swap providers like OpenAI, Anthropic, or Gemini with almost no code changes. It feels less like plugging in a model and more like giving your app a reliable coworker.
π @hackernewsgithubprojects
Laravel AI
Let your Laravel app talk to every AI provider without juggling ten different SDKs. This SDK gives you one simple, Laravel-friendly way to build smart agents that can use tools, remember conversations, and return clean structured data β plus generate images, audio, and embeddings. The coolest part: you can give an agent a job, let it think and act step by step, and even pause to ask you for approval before it does something risky. Then swap providers like OpenAI, Anthropic, or Gemini with almost no code changes. It feels less like plugging in a model and more like giving your app a reliable coworker.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ menuutux/maya-scene-io
Maya Scene IO
Maya Scene IO is the Blender add-on that finally lets Maya files breathe in Blender without needing Maya installed. It turns a Maya.ma or.mb file into a textured, skinned scene, right inside Blender, no Autodesk software required. The one vivid thing that makes it click: a character built in Maya can come through with its skeleton, skin weights and materials wired up, so it can be posed and deformed like it belongs there. The surprise is that the whole thing works offline, in Python, with no Maya open. It is not a perfect magic bridge, because rig logic does not transfer, but it gets you a clean, usable scene fast.
π @hackernewsgithubprojects
Maya Scene IO
Maya Scene IO is the Blender add-on that finally lets Maya files breathe in Blender without needing Maya installed. It turns a Maya.ma or.mb file into a textured, skinned scene, right inside Blender, no Autodesk software required. The one vivid thing that makes it click: a character built in Maya can come through with its skeleton, skin weights and materials wired up, so it can be posed and deformed like it belongs there. The surprise is that the whole thing works offline, in Python, with no Maya open. It is not a perfect magic bridge, because rig logic does not transfer, but it gets you a clean, usable scene fast.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ samuelwitbooi/fhotoroom-hdr-editor
Fhotoroom HDR
Fhotoroom HDR can shrink a JPEG by up to eighty percent while still looking sharp, and it does that right inside a Windows editor built for HDR work. Instead of juggling a RAW decoder, a tone-mapping tool, and a compressor, it brings the whole path into one place: decode, adjust, compress, export. It handles eighteen RAW formats from more than five hundred fifty camera models, keeps full detail while it works, and can pull in Lightroom files so you do not lose your catalog workflow. The catch is that it is free but not a magic fix: output quality still depends on your source files and settings, so keep backups.
π @hackernewsgithubprojects
Fhotoroom HDR
Fhotoroom HDR can shrink a JPEG by up to eighty percent while still looking sharp, and it does that right inside a Windows editor built for HDR work. Instead of juggling a RAW decoder, a tone-mapping tool, and a compressor, it brings the whole path into one place: decode, adjust, compress, export. It handles eighteen RAW formats from more than five hundred fifty camera models, keeps full detail while it works, and can pull in Lightroom files so you do not lose your catalog workflow. The catch is that it is free but not a magic fix: output quality still depends on your source files and settings, so keep backups.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ tencent/wemm-video
WeMM-Video
Watch a clip and hear it, too. That is the magic here: WeMM-Video reads text, images, audio, and video, and can even use a video's original soundtrack, then answers in plain text. Tencent built a bilingual family of models from two billion to thirty-five billion with three billion active parameters, trained on roughly 1.2 trillion multimodal tokens, with a fast instruct mode and a slower thinking mode that lays out the steps. The cool part is that the big version sets a new open-source mark for omni-modal video understanding and can describe what is on screen while summarizing speech in the same shot. Want the one vivid idea?
π @hackernewsgithubprojects
WeMM-Video
Watch a clip and hear it, too. That is the magic here: WeMM-Video reads text, images, audio, and video, and can even use a video's original soundtrack, then answers in plain text. Tencent built a bilingual family of models from two billion to thirty-five billion with three billion active parameters, trained on roughly 1.2 trillion multimodal tokens, with a fast instruct mode and a slower thinking mode that lays out the steps. The cool part is that the big version sets a new open-source mark for omni-modal video understanding and can describe what is on screen while summarizing speech in the same shot. Want the one vivid idea?
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ davanstrien/saturate
Saturate: The Self-Tuning Inference Pump
Saturate takes your messy data files and pushes them through any model server to return a clean, ready-to-use dataset. The coolest part is that it figures out the perfect speed on its own. Instead of guessing how many requests to send at once, it constantly adjusts its pace to match the server, just like your internet connection automatically speeds up when the traffic clears. You simply point it at your model and let it run. If you close the laptop or the job crashes, you can pick up right where you left off without losing any work.
π @hackernewsgithubprojects
Saturate: The Self-Tuning Inference Pump
Saturate takes your messy data files and pushes them through any model server to return a clean, ready-to-use dataset. The coolest part is that it figures out the perfect speed on its own. Instead of guessing how many requests to send at once, it constantly adjusts its pace to match the server, just like your internet connection automatically speeds up when the traffic clears. You simply point it at your model and let it run. If you close the laptop or the job crashes, you can pick up right where you left off without losing any work.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ software-mansion-labs/reverb-tts
Reverb-TTS: A Voice Clone That Runs on Your Phone
Most text-to-speech tools need a powerful GPU, but Reverb-TTS makes a new voice from just a few seconds of speech and runs it on a phone. Give it ten seconds of someone talking and their transcript, and it will speak any sentence in that voice, across seven languages, without any fine-tuning. It is built to be lightweight and edge-friendly, with a version that runs on an iPhone 17 faster than real time, while staying tiny enough to fit on your device. The surprising part is how fast and clear it is, even on a phone.
π @hackernewsgithubprojects
Reverb-TTS: A Voice Clone That Runs on Your Phone
Most text-to-speech tools need a powerful GPU, but Reverb-TTS makes a new voice from just a few seconds of speech and runs it on a phone. Give it ten seconds of someone talking and their transcript, and it will speak any sentence in that voice, across seven languages, without any fine-tuning. It is built to be lightweight and edge-friendly, with a version that runs on an iPhone 17 faster than real time, while staying tiny enough to fit on your device. The surprising part is how fast and clear it is, even on a phone.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ david-crespo/ah
ah: Talk to your coding agents from a browser
ah lets you open web pages that chat directly with the coding agent running in your terminal. It spins up a local server that connects your browser to your command line, so you can highlight a sentence in a document and ask for changes without leaving the page. The coolest part is that your agent can actually build these tools for you on the fly, creating a live dashboard that tracks your tasks and updates in real time. It turns your static text editor into an interactive workspace where code, comments, and results all live together.
π @hackernewsgithubprojects
ah: Talk to your coding agents from a browser
ah lets you open web pages that chat directly with the coding agent running in your terminal. It spins up a local server that connects your browser to your command line, so you can highlight a sentence in a document and ask for changes without leaving the page. The coolest part is that your agent can actually build these tools for you on the fly, creating a live dashboard that tracks your tasks and updates in real time. It turns your static text editor into an interactive workspace where code, comments, and results all live together.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ receitafederal/calculadora-rtc-regime-regular
Calculadora RTC Regime Regular
Here is the one thing that makes the calculadora-rtc-regime-regular feel genuinely different from most tax tools: it does not just tell you a total. It can build the official tax-related XML that Brazilian government systems expect, so a calculation can flow straight into a document instead of stopping at a number on a screen. That is the surprising part, because a normal calculator usually answers one question and then leaves you to do everything else by hand. This project turns tax logic into a usable pipeline.
π @hackernewsgithubprojects
Calculadora RTC Regime Regular
Here is the one thing that makes the calculadora-rtc-regime-regular feel genuinely different from most tax tools: it does not just tell you a total. It can build the official tax-related XML that Brazilian government systems expect, so a calculation can flow straight into a document instead of stopping at a number on a screen. That is the surprising part, because a normal calculator usually answers one question and then leaves you to do everything else by hand. This project turns tax logic into a usable pipeline.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ nestjs/agent-skills
NestJS agent-skills
The coding-agent helper that catches the NestJS mistakes that still pass tests. agent-skills gives Claude Code, Codex, Cursor and other coding agents the current NestJS APIs, first-party packages and real pitfalls, like sign-in checks that leak which emails exist or validation that lets null through. It is useful because the code can look correct while quietly being wrong, and the agent can load the right skill automatically when the task matches. The surprising part is that it teaches the agent what not to do, not just how to build.
π @hackernewsgithubprojects
NestJS agent-skills
The coding-agent helper that catches the NestJS mistakes that still pass tests. agent-skills gives Claude Code, Codex, Cursor and other coding agents the current NestJS APIs, first-party packages and real pitfalls, like sign-in checks that leak which emails exist or validation that lets null through. It is useful because the code can look correct while quietly being wrong, and the agent can load the right skill automatically when the task matches. The surprising part is that it teaches the agent what not to do, not just how to build.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ ssbullock/pcb-layout-llm
AI KiCad PCB Architectural Planner
Design a real circuit board before a human touches the layout tool. This project gives an AI the feel of an electronics engineer: it reads a compact description of a circuit board and hands back a clear plan for how the parts should be grouped and routed, so placement and routing can happen automatically instead of by hand. It was trained on many open-source KiCad boards, and in a blind test across 23 unseen boards, it produced valid plans every time and routed 83 percent of them completely, beating several large frontier models on the same test.
π @hackernewsgithubprojects
AI KiCad PCB Architectural Planner
Design a real circuit board before a human touches the layout tool. This project gives an AI the feel of an electronics engineer: it reads a compact description of a circuit board and hands back a clear plan for how the parts should be grouped and routed, so placement and routing can happen automatically instead of by hand. It was trained on many open-source KiCad boards, and in a blind test across 23 unseen boards, it produced valid plans every time and routed 83 percent of them completely, beating several large frontier models on the same test.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ finnmckenty/ntscrt
NTSCRT
NTSCRT is the little app that turns your photos and videos into broken, dusty, analog memory. It runs a real-feeling chain: a digital camera squeeze, the TV signal, a VHS tape and deck, dramatic TV glitches, then a CRT screen with scanlines and glow. The surprising part is that it does not just slap a filter on a clip; it simulates the whole chain, so the damage stacks the way real gear did, with tape flaws sitting in fixed spots and heads wearing differently from each other.
π @hackernewsgithubprojects
NTSCRT
NTSCRT is the little app that turns your photos and videos into broken, dusty, analog memory. It runs a real-feeling chain: a digital camera squeeze, the TV signal, a VHS tape and deck, dramatic TV glitches, then a CRT screen with scanlines and glow. The surprising part is that it does not just slap a filter on a clip; it simulates the whole chain, so the damage stacks the way real gear did, with tape flaws sitting in fixed spots and heads wearing differently from each other.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ lidija52/filmora-license-short-video-studio
Filmora License Short Video Studio
The surprising thing about Filmora License Short Video Studio is that it can fake access to a video editing tool for Windows without actually granting that access, and that makes it more of a risk than a shortcut. The repository presents itself as a compact studio for short-form editing, but the content points toward licensing and activation workarounds rather than a legitimate editing workflow. For a creator, the takeaway is simple: do not treat this as a trustworthy way to get software features.
π @hackernewsgithubprojects
Filmora License Short Video Studio
The surprising thing about Filmora License Short Video Studio is that it can fake access to a video editing tool for Windows without actually granting that access, and that makes it more of a risk than a shortcut. The repository presents itself as a compact studio for short-form editing, but the content points toward licensing and activation workarounds rather than a legitimate editing workflow. For a creator, the takeaway is simple: do not treat this as a trustworthy way to get software features.
π @hackernewsgithubprojects
This media is not supported in your browser
VIEW IN TELEGRAM
π¦ r0y00/xbox2windows
Xbox2Windows
Xbox2Windows can take pieces of Xbox Series X|S software and run them on a Windows PC. That means translating machine code and shaders instead of just launching a game, and it is impressive because the work is static, meaning it happens before any game actually runs. The surprise is that it already works: CPU functions, pixel shaders, BC7 textures and script states have been recompiled and checked against references. The project is interesting because it is not a game emulator but a careful experiment in rebuilding parts of a console runtime for Windows.
π @hackernewsgithubprojects
Xbox2Windows
Xbox2Windows can take pieces of Xbox Series X|S software and run them on a Windows PC. That means translating machine code and shaders instead of just launching a game, and it is impressive because the work is static, meaning it happens before any game actually runs. The surprise is that it already works: CPU functions, pixel shaders, BC7 textures and script states have been recompiled and checked against references. The project is interesting because it is not a game emulator but a careful experiment in rebuilding parts of a console runtime for Windows.
π @hackernewsgithubprojects