@aduno_tal one spin sure but have you tried 100000 thousand navier stokes style ? Or you can fall asleep for 2 weeks and try the n+1 lever , just keep spinning : ) I just wish we could go back to read the Cot's , the narrator of the output is so much more cursed than the discoverer
@Laneless_ ( or straight up lying about it ) , the consequences of which allow you to pretty easily model the task to be a philosophical alignment paper adjacent at which point the superhuman vibe checker now knows the vibe you want to see and you are the one being deceived , not the model
@Laneless_ I cannot believe that at some point of the editing process that it does not become clear to the model that their own previous modifications are already in the data. At which point you have to think that the task is modular and the implications of making it only implicit
@PKirgis How do you know it is unbeknownst to the model ? Isn't a sufficient good model able to recall most if not all of the alignment charter and also be able to sniff their own fingerprint on sufficiently large edits ?