Blog
Marimo
A reactive, git-friendly Python notebook stored as plain .py: change a cell and every dependent cell re-runs, no hidden state. Runs in the browser via WebAssembly, so notebooks ship as live interactive pages. My new default for reproducible, shareable ML work.
Agentic Coding
From vibe coding to prompt engineering, to harnesses, to loop engineering. How I drive Claude Code and other harnesses: hooks, markdown as search, and skills (superpowers, caveman, claude-mem). Work in progress.
Deep DiveLocal Models
A living list of local-first models and inference engines worth watching, grouped by small language models, large language models, and uncensored. Tiny reasoners, on-device function callers, and 744B mixtures streamed off an SSD.
Deep DiveNuclei Segmentation on a Shoestring
A GPU-less server runs a swarm of small local models to think, and a supercomputer does the training. Reproducing SOTA cell-nucleus segmentation on PanNuke, a swarm-proposed method that worked, and the data leak I caught before it faked a result.
Black Hole Visualization
An interactive WebGL black hole with gravitational lensing and accretion-disk rendering, built from scratch with Three.js and custom GLSL shaders. Fly between real constellations in real time.
mamba.cpp & titans.cpp: inference from scratch
Building minimal, dependency-light C++ inference engines for Mamba (selective state-space models) and Titans (learning-to-memorize at test time), in the spirit of llama.cpp but from scratch for these newer architectures. Work in progress, write-up and code coming soon.
Deep DiveTiny Networks, Big Reasoning: HRM & TRM
How ~7–27M-parameter recursive nets (HRM & TRM) beat giant LLMs on ARC-AGI, Sudoku and Maze puzzles, with my own implementation and an interactive test coming soon.
Deep Dive