Train a transformer right now

A real transformer — causal attention, two heads, layernorm — training on whatever text you paste, with the gradient written by hand and validated against finite differences. Within a minute it goes from spitting noise to mimicking your text's patterns. No libraries, no servers: one tab.