Writings

Undertow is a small synthetic benchmark experiment for hidden, stateful tool-use environments.

An exploratory diary of sparse-feedback experiments: judge sparsity, working memory, skill files, and the awkward places where the metrics disagree.

A note on mixed precision, high-precision parameter copies, and ZeRO.

How a single embedding table serves as both reader and writer in language models, with forward and backward passes and the split of gradients.