In a generation or two we'll have language models that understand aesthetics and accuracy (by being trained on token sequences annotated with aesthetic/accuracy

In a generation or two we'll have language models that understand aesthetics and accuracy (by being trained on token sequences annotated with aesthetic/accuracy scores), and we'll be able to ask the model to generate well written, factual answers by conditioning probabilities based on high aesthetic/accuracy scores.