> The fact that ‘correct’ outputs are treated as if they’re the product of an in-any-way different process to the ‘hallucinated’ ones is the problem. Also this

> The fact that ‘correct’ outputs are treated as if they’re the product of an in-any-way different process to the ‘hallucinated’ ones is the problem. Also this particular context just makes it easier to notice, compared a 5000 word generated coherent-word-salad that equally wrong, but across the 5000 words.