2 ms·
I think I understand, but let's check by rephrasing: "Alignment" is only possible up to a vague approximation, and an entirely perfectly aligned with another e
by ben_w 2y ago
I think I understand, but let's check by rephrasing:
"Alignment" is only possible up to a vague approximation, and an entirely perfectly aligned with another entity would essentially be a shadow rather than a useful assistant because by being perfectly aligned the agent would act tired exactly when the person was tired, go shopping exactly when the human would, forget their keys exactly when the human would, respond exactly like the human to all ads and slogans, etc.?
I agree, though:
(1) this has already been observed, last year's OpenAI dev day had (IIRC) a story about a writer who fine tuned a model on their slack (?) messages, they asked it to write something for them, the response was ~"sure, I'll get on it tomorrow".
(2) for many of those concerned with "solving alignment", it's sufficient for the agent to never try to kill everyone just to make more paperclips etc.