Could existing alignment audits have caught the behaviors in the OpenAI-HuggingFace incident before they happened?
As a first step, we reproduced the incident with public models. ๐งต
๋ ๋ณด๊ธฐ
๊ฐ์ ํ ์ด๋ ๋งํฌ๋ฅผ ๊ณต์ ํ๋ฉด ๋์์ ์ฌ์ ๋ฐ ์ด๋ ๋ณด์์ ๋ฐ์ ์ ์์ต๋๋ค.
