Writing
Most of these are reproductions of new model research or post-mortems of two competitions I lost. When a claim turns out wrong, I correct the post and mark the change.
The Work You Don't Do: Losing Two Optimization Competitions the Same Way
3,768 words, 17 min
I named the question I failed to ask, then entered another competition and failed to ask it again. What actually wins these races, and why I keep missing it.
2 competitions entered , then 2 lost the same way
Competing in an open ML-optimization competition on hardware that could not run the ranked benchmark, and why measurement bandwidth decided everything.
8 submitted , then 0 promoted
Same Weights, Same Prompt, Opposite Results - Why Local Tool Calling Breaks
3,068 words, 14 min
I built willitcall, a conformance suite and public matrix for tool calling on local models. The findings that held up, and the one I retracted.
Blamed Ollama's parser , then decoding differs, claim retracted
A standalone kit that installs a reliable orchestrator-delegates-to-executor workflow into Claude Code - and the dogfooding loop that wrote it
8 tasks delegated , then 2 failed, both my spec's fault
I Tried to Break Google's New Tabular Foundation Model. Then I Fixed It.
2,875 words, 13 min
An independent, reproducible evaluation of TabFM - what held up, what did not, and the bug fix that got merged into Google's repo
Won on 10 datasets , then demoted 2 of my own wins to ties
TurboQuant on an M1 Pro - needle retrieval from 0% to 100% at 16K on MLX, and four bugs fixed in two llama.cpp forks
Called it a bug fix , then variance reduction, corrected