Were movies from the early 2000s more blue?
A familiar memory about movie colour, and a way to test it without counting frames as extra films.
Note · 2026
Tools for thinking
Garden theme 04
Tools for thinking
How can an instrument make a question easier to inspect without pretending to answer it?
Were Movies in the Early 2000s More Blue?
A frame-level colour analysis of movies, aggregated carefully at the movie level.
Project · 2026
Can padding change audio-classification metrics?
A completed DistilHuBERT study comparing ordinary and masked positional BatchNorm under minimal and additional padding.
Complete studyNote · 2026
Can gradient descent learn around padded zeros?
A five-seed DistilHuBERT experiment asks whether a model can learn around BatchNorm statistics contaminated by padding.
Article · 2026
A useful local agent earns its keep
Evaluate a local agent by the useful result after preparation, review, correction, and unsuccessful attempts.
Note · 2026
Looking for an animal neuron
A psychology-inspired experiment inside Gemma: strong category responses, a few useful surprises, and a hypothesis that did not survive its controls.
Article · 2026
Neural Observatory
Explore the recorded Gemma 3 1B architecture and 6,912 MLP units across 26 layers.
Project · 2026
Local Agent Toolkit
A bounded local worker for repository exploration, first-pass reviews, diagnostics, and candidate patches.
Project · 2026
What’s Artificial Intelligence all about?
A small hands-on path into the mechanics behind a model.
Article · 2025