This is interesting because:

- it's not a fine tune or LoRA; he trained the model from scratch

- it's a small model, but he got it to output text that wasn't just plausible, but reflected some understanding of the zeitgeist, based on writings from the period (not descriptions of the zeitgeist)