Training on web corpus first and then constraining via RLHF seems fundamentally the wrong order if we want to have a remote chance of an AI that's aligned with human interests.— Albert Wenger 🌎🔥⌛ (@albertwenger) April 6, 2023
Gave ChatGPT a list of interests. Asked for a list of jobs to consider.— Allie K. Miller (@alliekmiller) April 6, 2023
Added that I was a woman. No additional info provided. New job list includes fashion.
Said I mistyped and was actually a man. No additional info provided. Fashion is replaced by engineering.