Yes, it submits lots of varied prompts that get refused, and then lots of varied prompts that don't get refused, then iteratively edits weights so those two groups end up in roughly the same latent space.
replyI wonder, if we could accurately resolve discrete neural signals in the human brain, if a similar process would work.
reply