Hacker News Reader

These are all pretty cool. What determines how it could "improve" generated words? Perhaps a larger, "pleasant"-words-only corpus?

Larger corpus, deeper model, better scoring function on models, more models