AI systems, Local LLMs, Agents, GPUs I test models, read code, and post the numbers.
Seoul, KR
Joined September 2013
- Tweets2.7K
- Following156
- Followers1.8K
- Likes6.5K
Pinned Tweet
I happily picked up two used Nvidia Tesla P40 GPUs.
I knew the P40 was old and ridiculously overbuilt.
What I did not expect was how long the GPUs would become once the coolers were attached.
Tried putting them in the case.
Nope.
The cooler end would not fit.
I thought about it for about 10 seconds.
Turns out the bottleneck was not the human.
It was the case.
So I grabbed a hammer and pliers and destroyed the front section of the case.
After that both P40 GPUs fit perfectly.
Put the cover back on and you cannot even tell from the outside.
Apparently hardware fits if you hit it hard enough.
Good experience.
TinyGo
Take Go to microcontrollers, browsers, and edge devices
TinyGo compiles Go programs using LLVM, with a focus on very small binaries and avoiding unused features.
It supports over 150 microcontroller boards. You can also target browsers through WebAssembly and servers or edge devices through WASI.
WebAssembly programs can run in Fastly Compute, Fermyon Spin, wazero, and other WebAssembly runtimes.
github.com/tinygo-org/tinygo
graphify
Turn your codebase into a graph your coding agent queries instead of grepping
Type /graphify in Claude Code, Codex or Cursor. It maps code, docs, PDFs and images into a knowledge graph.
Code is parsed locally with tree-sitter, with no LLM and nothing leaving your machine. Every edge is tagged as extracted from the source or inferred, so you can see what is certain.
It is a real graph, not a vector index. You can ask a question, trace the path between two parts, or open an interactive graph.html of the whole project.
github.com/Graphify-Labs/gra…
React Bits
215+ animated React components and backgrounds you can try before adding
Every component has a live demo with controls, so you can adjust it in the browser before adding it to your project.
Choose JavaScript or TypeScript, with CSS or Tailwind styling. Add components through shadcn or jsrepo, or copy the files from the component's Code tab.
Through the shadcn MCP server, Claude Code, Cursor, Codex, VS Code, and Antigravity can find and add components.
github.com/DavidHDev/react-b…
Trowel
Type a build in your terminal and watch it appear in Minecraft as a blueprint first
You type a request like a watchtower on the hill, and Trowel reads the blocks around you, the biome and what you are looking at. The build shows up as white outlines only you can see. Say yes and it fills in, or take the whole turn back with one undo.
It matches nearby materials, so a tower next to a spruce house comes out in spruce and stone. It can also write datapacks for recipes, advancements and loot tables, then reload and fix what broke.
For survival, it lists every block you need and can export a Litematica file to build by hand.
github.com/Straitbrashadow/m…
ManimGX
Renders math and 3D animation videos 90.6x faster than ManimCE on an Apple M4 Pro
ManimGX is an animation engine that takes Python written with Manim CE's API and renders it on a Rust and wgpu core. It is built for coding agents to write scenes.
Test: five scenes, 36.2 seconds of video at 1080p60, on an Apple M4 Pro. Time from launch to finished MP4, fastest encoding preset for every engine, lower is better.
ManimGX: 3.08 s total
90.6x faster than ManimCE
11.5x faster than ManimGL
18.7x faster than Blender Workbench
It mixes 2D and 3D in one scene, writes math in LaTeX or Typst with no LaTeX install, and can sync animation to speech. It also runs in the browser through Pyodide and WebGPU.
github.com/academa-labs/mani…
Max Headbox
A voice agent that lives on a Raspberry Pi 5 and never calls a cloud model
Max Headbox listens for a wake word, transcribes what you say, and lets a small local model pick a tool to run on the Pi. Nothing leaves the device.
It runs gemma3:1b and gemma4:e2b through Ollama and needs about 6GB free for the models. Tested boards: 8GB and 16GB. Vosk handles the wake word and faster-whisper does the transcription.
A tool is one JavaScript file in src/tools with a name, parameters, a description and the code to run. Mark a tool dangerous: true and the agent asks you YES or NO before it acts.
The backend is now Express.js, so Node 22, Python 3 and Ollama are all you need.
github.com/syxanash/maxheadb…
whiteboard-animation-vi
Turn any illustration into a whiteboard drawing video, characters colored first
Upload one image and it draws the line art, colors the characters, then fills the background, with real photographic hands moving the pen.
An anime segmentation model finds the foreground so characters appear early. Colors are grouped along the ink lines, and the result keeps the original colors and shading.
The hand route is optimized to cut travel between strokes, and the samples render 8 seconds at 50 FPS, up to 1280x720.
Runs locally in a Gradio app with CUDA and NVENC, or on CPU.
github.com/letuandat131/whit…
Machine Skill
Describe a machine in one line and Claude builds a working 3D simulation of it
Machine Skill is an open source Claude Code skill. Claude researches the real machine, writes down its dimensions and the physics behind it, and builds a photoreal, interactive simulation in one HTML file with three.js.
Every number on screen is computed by a small solver. The turbofan example reports about 340 kN at take off, 83 % of it from the bypass, across 3,597 parts. The rocket engine example gives 2,462 kN and Isp 321 s at 330 bar.
You can drag, cut open, explode and break each machine. Drop the tower level on the refinery pump and it cavitates.
One command records a 28 s film of the build for posting.
github.com/AirsupHQ/machine-…
Piskel
Make game sprites and pixel art animations in your browser and export animated GIFs
Piskel is a simple sprite editor that runs on the web. It is the editor behind piskelapp.com.
It is built with plain JavaScript, HTML and CSS. Animated GIFs are generated in the browser with web workers, and GIFs can also be imported and edited.
You can use it in the latest Chrome, Firefox and Edge. Brave needs canvas fingerprinting protection turned off. There is no mobile support, but offline desktop builds are available.
github.com/piskelapp/piskel
arduino-esp32
Use Arduino on ESP32, including inside ESP-IDF projects
arduino-esp32 is the Arduino core for the ESP32 family of chips. You can also use Arduino as a component in an ESP-IDF project.
The docs explain API compatibility with ESP8266 and Arduino-CORE (Arduino.cc). CI tests cover external libraries used with the core.
EspExceptionDecoder turns exception output into meaningful call traces. Installation guides cover Windows, Linux, and macOS.
github.com/espressif/arduino…
browser-use
An open source browser agent that clicks through websites for you, with any LLM
browser-use is an open source browser agent. You describe a task in plain language, and the agent opens a browser, works through the pages and returns the answer.
It runs locally with any LLM, or on Browser Use Cloud. You can use the hosted cloud agent, a command line tool for your own browser tasks, or a Python library inside your own code.
The same code runs on a local Chrome or a cloud browser. Cloud browsers cost $0.02 per browser hour plus traffic, and add proxies, CAPTCHA solving and a live view.
It is for developers who want an agent to fill forms, look things up and click through sites.
github.com/browser-use/brows…
LDtk
A modern open source 2D level editor that puts ease of use first
LDtk, the Level Designer Toolkit, is a modern, efficient and open source 2D level editor with a strong focus on user friendliness.
Download the latest version from ldtk.io. It can also run inside a Hide editor tab, and its renderer can run on its own in NW.js without the Electron main process.
New projects start with the Endesga32 palette by Endesga. A Haxe API is on GitHub for loading levels in your game.
github.com/deepnight/ldtk
ProtoMotions 3
NVIDIA's GPU framework for physics based digital humans and humanoid robots
ProtoMotions trains simulated humans to move with physics, from motion capture data. The same setup trains humanoid robots.
It runs on IsaacLab, Newton, IsaacGym and MuJoCo. MaskedMimic policies can finish a motion from partial goals instead of copying one clip.
On Oct 28, 2026 the team released Metatron, a whole body policy for the Unitree G1.
github.com/NVlabs/ProtoMotio…
