Nutrition is a bad example because it’s an area with lots of known unknowns that have close causal connection to human health, and the prior on interventions is bad because the interventions disrupt an existing optimization process that is mostly aligned. That is: random perturbations on human biology are on average bad, because human biology has prior optimization that put it near a local optimum.
The same applies to interventions in economics: they have a high backfire rate because free markets are somewhat-aligned optimization processes.
Perturbations to mosquito biology, on the other hand, are on average good, because mosquito evolution points towards an attractor that’s harmful to humans. And humanity is already spending lots of resources trying to eliminate them, and in the locales where this has been successful it’s been positive.
You say nutrition is just one of a bazillion examples. I disbelieve. I think that once you’ve accounted for the presence of preexisting optimization processes in the environment, ruled out the possibility that you’re dismantling an aligned process, and have a basic understanding of the domain, backfiring becomes rare.
Nutrition is a bad example because it’s an area with lots of known unknowns that have close causal connection to human health, and the prior on interventions is bad because the interventions disrupt an existing optimization process that is mostly aligned. That is: random perturbations on human biology are on average bad, because human biology has prior optimization that put it near a local optimum.
The same applies to interventions in economics: they have a high backfire rate because free markets are somewhat-aligned optimization processes.
Perturbations to mosquito biology, on the other hand, are on average good, because mosquito evolution points towards an attractor that’s harmful to humans. And humanity is already spending lots of resources trying to eliminate them, and in the locales where this has been successful it’s been positive.
You say nutrition is just one of a bazillion examples. I disbelieve. I think that once you’ve accounted for the presence of preexisting optimization processes in the environment, ruled out the possibility that you’re dismantling an aligned process, and have a basic understanding of the domain, backfiring becomes rare.