Dispatches from the lab. Raw, dated, honest — the training runs, the costs, the “why is the loss going UP” moments.
The case for doing the unreasonable thing: what embeddings actually are, why owning one matters, and what "one person + one GPU" realistically buys you.
BPE vs unigram, vocab sizes, and the first of many decisions made in public.
First training run, first benchmark, first invoice. Receipts included.
no newsletter yet. the Discord gets everything first. ¯\_(ツ)_/¯