research: ChatGPT can emit sequences memorized from its training data via a new "divergence attack" that prompts ChatGPT to repeat a word numerous times (Alex Ivanovs/Stack Diary)

Alex Ivanovs / Stack Diary:

research: ChatGPT can emit sequences memorized from its training data via a new “divergence attack” that prompts ChatGPT to repeat a word numerous times  —  Large language models, appreciate ChatGPT, are trained on vast amounts of text data from books, websites, and other sources.