“Also that IQ tests and tests like the SAT enjoy significant predictive validity.”
Since I started to learn machine learning, I’ve come to appreciate how a low bar that really is. Making a test that has “predictive validity” for any of the usual touted outcomes, given a reasonable amount of data, is trivial. There’s this folk wisdom in IQ fan circles that it’s impossible to make a better predictive model than IQ — that is utterly wrong. Add in economic status. Add in PE grades. Add in the multitude of tests and tasks the IQ fans have thrown out for not being “g-loaded” enough (which is the same as not providing support for their model). Even without that, you can certainly train a better model if you don’t restrict yourself to early 20th century methods.
But there’s also unsupervised learning, where you try to learn a model which represents the variation in the data well. Although this will rarely be optimal for any specific predictive task, the hope is that it will provide a “good enough” model for lots of different tasks, and that it can give a good starting point for exploration of the variation. IQ doesn’t do well there either, as you’d expect from a one factor model.
The one thing IQ has going for it is lots of historical data. But even that should be treated with caution — it’s dubious to say that what Terman’s tests measured is the same as what today’s tests measure.