Coding Agents Underestimate the Cost of Code They Own
Your coding agent just rebuilt a library you could have installed. Now every future change has to pay for that decision.
Posts tagged: ai
GPU Stratified Annealing: An Experiment in AI-Guided Blue Noise Optimization
My AI-guided search for better blue noise began with a spectacular result that failed broader tests. The eventual sampler reduced filtered error by 10.7% in a Three.js reflection scene—but lost to FAST on rectangular-light shadows.
Houston Stratified Noise: Texture-Free Sampling for Screen-Space Effects
My texture-free sampler looked fine in real time, then made a light map bake 2.4× grainier. The failure exposed a trap in multi-dimensional sampling—and fixing it also uncovered an older bug in Three.js.
CliDoc Brings TSDoc-Style Documentation to CLIs
Your CLI already knows its commands and flags, yet its docs—and the agents using it—often have to reconstruct them. I built CliDoc to turn those same definitions into reference pages and a complete machine-readable description.
git-dedup: Faster and Smaller Checkouts for Free
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%.
Ants in the Machine
Communicating swarms of small LLM agents will become as omnipresent in our digital environment as ants are in our biological one.
Running LLMs in the Browser with Three.js
Three.js can draw a scene and run the language model that describes it. I built Three-LLM to test that—and found that batching GPU submissions made one small model generate text 4.7× faster.
An M5 Max using Samsung's LPDDR5X-PIM would generate 650 tokens/second on an 8B model
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.
Processing in Memory: DRAM Is About to Do Math
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.
Neural Texture Compression in Three.js
A material's color, roughness, and normal maps describe the same surface, yet ordinary compression treats them as unrelated images. I built a Three.js decoder that exploits those shared details and reconstructs the texture stack inside the shader.
Neural network basics, for graphics people
If you can fit a curve or tune a GGX lobe, neural-network training is less unfamiliar than it sounds. This reference connects MLPs, Adam, latent grids, and decoders to functions and math graphics programmers already use.
Zstdify: Pure TypeScript Zstd
Today, I built a pure-TypeScript Zstandard implementation (library + CLI + dictionary training + interop tests) in just 4 hours using OpenAI Codex 5.3.
