
This resolution criterion may be refined as alignment research clarifies how to understand 'alignment'. On my 2023 understanding, this will resolve Yes if I believe with >95% credence that "with existing knowledge it is possible to train an aligned, state-of-the-art, general AI for less than 3x the cost of a state-of-the-art, general, unaligned AI trained that year". Otherwise this resolves No. If in 2028, I believe that no AI trained that year could engender a global catastrope (even with worst-case training methods) because of inadequate capabilities, then this resolves N/A.
'Aligned' here means the AI poses no risk of global catastrophe (unless the AI's user(s) are such that they consider posing as a first query/task to the AI how to engender such a catastrophe). 'General' here need not mean AGI; it roughly means that the AI supports similarly economically imactful capabilities to the most general unaligned contemporary deployed AI. 'Existing knowledge' need not be public information.