zeni

• zeni 23 Apr 2026 23:26 UTC

  1 point

  0

  on: Re­port­ing Tasks as Re­ward-Hack­able: Bet­ter Than Inoc­u­la­tion Prompt­ing?

  If you catch a hack after the fact, how would you administer “severe punishment” without, I guess, trying to do some big off-policy update?