I wrote back to Abram Demski regarding his comments from fourteen months before: on further thought, he was right. Even granting my point that evolution didn't figure out how to track probability and utility separately, as Abram had pointed out, the _fact_ that it didn't meant that not tracking it could be an effective AI design. Just because evolution takes shortcuts that human engineers wouldn't didn't mean shortcuts are "wrong". (Rather, there are laws governing which kinds of shortcuts _work_.)
Abram was also right that it would be weird if reflective coherence was somehow impossible: the AI shouldn't have to fundamentally reason differently about "rewriting code in some 'external' program" and "rewriting 'its own' code." In that light, it made sense to regard "have accurate beliefs" as _merely_ a convergent instrumental subgoal, rather than what rationality is about—as sacrilegious as that felt to type.
And yet, somehow, "have accurate beliefs" seemed _more fundamental_ than other convergent instrumental subgoals like "seek power and resources". Could this be made precise? As a stab in the dark, was it possible that the [theorems on the ubiquity of power-seeking](https://www.lesswrong.com/posts/6DuJxY8X45Sco4bS2/seeking-power-is-often-robustly-instrumental-in-mdps) might generalize to a similar conclusion about "accuracy-seeking"? If it _didn't_, the reason why it didn't might explain why accuracy seems more fundamental.
I wrote back to Abram Demski regarding his comments from fourteen months before: on further thought, he was right. Even granting my point that evolution didn't figure out how to track probability and utility separately, as Abram had pointed out, the _fact_ that it didn't meant that not tracking it could be an effective AI design. Just because evolution takes shortcuts that human engineers wouldn't didn't mean shortcuts are "wrong". (Rather, there are laws governing which kinds of shortcuts _work_.)
Abram was also right that it would be weird if reflective coherence was somehow impossible: the AI shouldn't have to fundamentally reason differently about "rewriting code in some 'external' program" and "rewriting 'its own' code." In that light, it made sense to regard "have accurate beliefs" as _merely_ a convergent instrumental subgoal, rather than what rationality is about—as sacrilegious as that felt to type.
And yet, somehow, "have accurate beliefs" seemed _more fundamental_ than other convergent instrumental subgoals like "seek power and resources". Could this be made precise? As a stab in the dark, was it possible that the [theorems on the ubiquity of power-seeking](https://www.lesswrong.com/posts/6DuJxY8X45Sco4bS2/seeking-power-is-often-robustly-instrumental-in-mdps) might generalize to a similar conclusion about "accuracy-seeking"? If it _didn't_, the reason why it didn't might explain why accuracy seems more fundamental.