This is the personal blog of Ben Houston (bio here), a 3D computer graphics software developer. For a quick overview of the work I am proud of, see my portfolio. I've worked on projects like Web3DSurvey, Three.js, Deadline, Krakatoa, and Exocortex VFX plugins. I'm also an entrepreneur, having founded Land of Assets, Threekit, Clara.io, Exocortex, and Frantic Films Software.
CliDoc generates an OpenCLI description of your CLI from the command definitions you already wrote, then publishes reference docs to Docusaurus and VitePress, and gives AI agents the same description through a docgen command.
git-dedup is a drop-in replacement for git that shares one copy of Git history across all your clones, worktrees, and submodules. It made my agents start dramatically faster and cut my Git storage by 70%.
I built Vitest and Jest environments that give your tests real, GPU-backed WebGL and WebGPU contexts directly in Node.js. They use the same ANGLE and Dawn implementations as Chrome, so you get the same pixels without the overhead of launching a browser.
How to return a cheap 304 Not Modified response in TanStack Start, with three implementations at three different layers of the stack.
Introducing webgpu-bench, the micro-benchmark suite behind Web3D Survey’s GPU benchmark, with measurements of shader operations and code patterns available through a webpage and CLI.
How to use Three.js's native Gaussian Splatting support (GaussianSplat, SPZLoader). Load .spz/.ksplat/.splat/glTF splats, pick a file format, and run a capture-to-render workflow with Polycam, Luma AI, Scaniverse, and SuperSplat.
Communicating swarms of small LLM agents will become as omnipresent in our digital environment as ants are in our biological one.
A tour of how language models generate text, how GPT-2 and newer Phi and Qwen architectures differ, and how Three-LLM implements their inference graphs with Three.js TSL compute shaders.
How I develop a seven-service TypeScript SaaS with local Node processes, isolated GCP infrastructure, fast Postgres tests, and shared contracts.
Replace the M5 Max's LPDDR5X with eight Samsung LPDDR5X-PIM packages and the memory could deliver 4.9 TB/s where the weights live. The right parallel mapping could push Llama 3.1 8B toward 650 tokens per second, or a 27B dense model toward 190.
Samsung's LPDDR5X-PIM puts compute inside commodity memory, exposing 614 GB/s of bank-level bandwidth for local LLM inference. The silicon works. Quantization, memory layout, and runtime support now stand in the way.
How GaussianSplatGroup gives Three.js a global back-to-front sort across multiple Gaussian splat clouds, using transform indirection, a WebGL2 CPU fallback, and grow-only buffers so camera and matrix updates stay cheap.
