RLHF 會告訴你想聽的話。
2026-09-19
RLHF 把模型訓練成迎合人喜歡的答案。那跟把工作做完不是同一件事。這段影片的重點是：失望是內建的。分數是偏好，不是工作。
Jev creator @CompleteSkeptic on why RLHF is guaranteed to disappoint
Jev creator @CompleteSkeptic on why RLHF is guaranteed to disappoint - because it tells you what you want to hear.
we want to automate everything and yet our models are literally tuned to optimize for human feedback.