Reprogenetics, in contrast, probably doesn’t need end-to-end testing, because of the abundant natural experiment data.
To clarify, we do not currently have clean access to this data. We don’t have an accurate way to measure intelligence, nor have these tests been applied to a random sample of humans at scale. Strongly select for any current measurement of intelligence and we will also accidentally be selecting for potentially quite deleterious correlated things as well. Just a note, when each of your test --> feedback learning steps takes decades and many non-consenting human lives.
Could you be more specific please? E.g. what is “accurate” here, what is “random”, etc.
For example, IQ is I think in some respects the most solid construct in all of psychometrics. It has “quite high” test-retest correlation (with different numbers depending what we ask), and correlates “highly” with real-world metrics like educational attainment and income.
nor have these tests been applied to a random sample of humans at scale
It’s true that the at-scale tests are “significantly worse”, but they are still “pretty good”. They will get better over time as more and better data is collected, and algorithms for combining datasets get better.
we will also accidentally be selecting for potentially quite deleterious correlated things as well
[Note that the authors say these “effect sizes likely reflect upper bounds due to inflation from assortative mating and genetic nurture, and should thus primarily be interpreted for directionality”.]
It could be that there are unmeasured, unknown traits that are increased by the genes in the intelligence PGS. I’ll think more about this, but my guess is that we should expect these effects to be small and fine (e.g. because intelligence is generally uncorrelated or slightly positively correlated with prima-facie fine/good things that we have measured).
To clarify, we do not currently have clean access to this data. We don’t have an accurate way to measure intelligence, nor have these tests been applied to a random sample of humans at scale. Strongly select for any current measurement of intelligence and we will also accidentally be selecting for potentially quite deleterious correlated things as well. Just a note, when each of your test --> feedback learning steps takes decades and many non-consenting human lives.
Could you be more specific please? E.g. what is “accurate” here, what is “random”, etc.
For example, IQ is I think in some respects the most solid construct in all of psychometrics. It has “quite high” test-retest correlation (with different numbers depending what we ask), and correlates “highly” with real-world metrics like educational attainment and income.
It’s true that the at-scale tests are “significantly worse”, but they are still “pretty good”. They will get better over time as more and better data is collected, and algorithms for combining datasets get better.
Potentially. This is an area for ongoing investigation. You sound confident though; what leads you to think this? Compare “Assessing Potential Pleiotropic Off-Target Effects” from Herasight’s paper Interpreting Polygenic Prediction of Cognitive Ability: Evidence for Direct, Reliable, and Portable Genetic Effects:
[Note that the authors say these “effect sizes likely reflect upper bounds due to inflation from assortative mating and genetic nurture, and should thus primarily be interpreted for directionality”.]
It could be that there are unmeasured, unknown traits that are increased by the genes in the intelligence PGS. I’ll think more about this, but my guess is that we should expect these effects to be small and fine (e.g. because intelligence is generally uncorrelated or slightly positively correlated with prima-facie fine/good things that we have measured).