7 projects · 4 written up
Things I have built, mostly to understand something rather than to ship it. Four of them have an entry on the log, which is usually the more honest account: what I expected, what actually happened, and which part turned out to matter. The rest are just the code, for now.
A terminal-native multi-agent research tool. It splits one question into parallel web-searching subagents, supervises them with Sotis, and merges what comes back into an exportable report, all inside a keyboard-driven TUI.
Watches an LLM agent while it is running and steps in when it starts to spiral, instead of reading the wreckage afterwards.
Reproducing gwern's LLM-catapult hypothesis on a laptop GPU: can a high learning rate and heavy weight decay push a network through a memorisation-to-algorithm phase transition? Grokking reproduced on modular addition, with weight decay turning out to be the load-bearing knob.
Read the log entry →A signed-distance-field ray marcher in about 170 lines of plain Java. No engine, no shader language, no graphics library, plus anti-aliasing and soft shadows.
Read the log entry →Four continuous chaotic systems and the Clifford map, rendered from one framework, including the one that refused to fit it.
Read the log entry →A Mandelbrot renderer with no cache and no database. It rewrites its own source file to remember what it has already drawn.
Read the log entry →A self-organising map built to behave less like a static grid of data and more like something alive.
Everything else lives on GitHub.