The writeup includes example text where the algorithm is fed a sentence starting about George Washington and within half a sentence or so goes unhinged and starts praising Trump...

Also, a reminder to folks that this model is not conversationally trained and won't behave like ChatGPT; it cannot take directions.