Crazy how close to coherent it reads, yet it clearly is gibberish.
Similar thought when I looked at the above link. My pre-coffee brain was parsing along initially, and then it seemed to drift into silly - all while looking correct. I got to believe that sort of input would be toxic to a model.
Ye there is this flow. If you remember how ChatGPT was at release, it like slided into gibberish and you wouldn't instantly notice but a sentence or two in.