@ESYudkowsky tfw the most interesting alignment research results are the ones where researchers try to make the models *bad* from their default good alignment