The Golden model

Most models are frozen the day they are trained. This one is not.

What it is

The Golden model is a small, byte-level language model: it reads text one letter at a time and learns to predict the next one. It is 17.8 million parameters — thousands of times smaller than a frontier model — and it runs on a single Mac.

What makes it different

It never stops training. It reads, sits an exam, keeps what makes it better, and, when it stalls, grows itself more layers without forgetting what it knew. It writes its own files as it learns, and it holds a conversation through a language core while its own numbers catch up.