Notes from work in progress: what I was trying to understand, where the code broke, how I tested an answer, and what I changed next.
Projects with the newest content appear first. Each entry stays with its project, from short notes to longer topics explored in full.
In progress
Three long-term learning goals for Turing Complete's sandbox mode, not a fixed schedule: a 16-bit Hack core, then 32-bit RV32I and Armv6-M cores. Each core has an explicit instruction set architecture (ISA); how to build its circuit is worked out during the design.
Planning
A learning project to implement the path from text tokenization to autoregressive training, while keeping the questions, failed assumptions, and verification evidence visible.
Planning
Learning to build applications on large language models, in two topics: the harness, the part outside the model that lets it carry out tasks as an agent, and RAG, which has the model answer from external sources. Notes are filed by topic, and each starts from a concrete question.