Research preprints, long-form notes, translations, and practical engineering material.
A survey of neural network weight generators: hypernetworks, seeds, codebooks, structured transforms, fractals, and recursion. Storage, memory, and compute budgets, with selected reproductions on CPU and a MacBook.
Seed thresholds for heterogeneous bootstrap percolation with thresholds one and two, comparing prescribed-degree and independent-edge random graphs as the mean degree grows.
Two sets of Headroom experiments: 80 agent runs, token usage, task success, and the gap between request compression and savings over an entire task.
A follow-up to the SKILL.state experiments: total and uncached input, cached-token share, output, and full success across Native, Paper, V2, and V3.
An experimental comparison of SKILL.state memory with the standard conversation history of OpenCode and Codex across 820 runs of five software-development tasks.
How understanding, control, and task decomposition turn AI assistants into an engineering tool rather than a productivity gamble.
A citation-network approach to retrieving legally significant documents from a corpus of 31,658 decisions of the Constitutional Court of the Russian Federation.