I am not surprised that gold background is a undesirable trait. however this is how we get high side-effects for women in drugs sold at stores, because testers prefer male over female. If humans in the wild have a 20% trait rate and your sample has 1% or 0% that is going to lead in a bad result in its own way. Having a WEIRD sample is not particularly representative.
If you have a discipline that supports multiple frameworks and recruit on resonance with a particular framework then the result tells less about the frameworks properties. For example one could try to provde that chess is an endurance game of bothering to check enough positions and reqruit based on stamina in order to “prove” it is not a game of intellect.
I remember when balancing away dive was a talking point. Then a lot of the teams were squamish in scrimming other strategies. If you need to redo the whole strategy stack instead of just adjusting the top layers then teams will eventually do it but it can take long while.
If you tell a high rank player to push they will know to still reftrain from being mindlessly suicidal, to not push all the way throught spawn etc. If you desribe somethings color in grue and bleen if helps if the communication reciever has existing support for those concepts. Even if there is no explicit culture sharing the learning curve could provide a way for “on the onset” some fundamentals to be evident and then when those are taken into account then more fine-graded concepts can make sense. But part of the point is that the incentive gradient to make the distinction doesn’t exist at all stages. This can be seen as an aspect to the “smiley face maximiser” error state of aligment problem, the defintions and concepts that humans actually use don’t exist in a neat context-free way. Telling a human to go “make people smile” result in sensible action while a literal minded Ai will tile things destructively with inapproriate patterns.
I am not surprised that gold background is a undesirable trait. however this is how we get high side-effects for women in drugs sold at stores, because testers prefer male over female. If humans in the wild have a 20% trait rate and your sample has 1% or 0% that is going to lead in a bad result in its own way. Having a WEIRD sample is not particularly representative.
If you have a discipline that supports multiple frameworks and recruit on resonance with a particular framework then the result tells less about the frameworks properties. For example one could try to provde that chess is an endurance game of bothering to check enough positions and reqruit based on stamina in order to “prove” it is not a game of intellect.
I remember when balancing away dive was a talking point. Then a lot of the teams were squamish in scrimming other strategies. If you need to redo the whole strategy stack instead of just adjusting the top layers then teams will eventually do it but it can take long while.
If you tell a high rank player to push they will know to still reftrain from being mindlessly suicidal, to not push all the way throught spawn etc. If you desribe somethings color in grue and bleen if helps if the communication reciever has existing support for those concepts. Even if there is no explicit culture sharing the learning curve could provide a way for “on the onset” some fundamentals to be evident and then when those are taken into account then more fine-graded concepts can make sense. But part of the point is that the incentive gradient to make the distinction doesn’t exist at all stages. This can be seen as an aspect to the “smiley face maximiser” error state of aligment problem, the defintions and concepts that humans actually use don’t exist in a neat context-free way. Telling a human to go “make people smile” result in sensible action while a literal minded Ai will tile things destructively with inapproriate patterns.