When the U.S. air force discovered the flaw of averages

Perhaps for a naive recommendation system, but quality ones don't do this.

For example if you employ 1-nearest neighbors, you will actually return a sample that is guaranteed to reproduce another user's preferences. You could of course sample over more neighbors, smoothing results, at the cost of more lookups and at some point in the limit you do see reduced performance due to too many points becoming similar and having this average of everyone being bad for everyone sort of thing. Most ML approaches account for this and will only look at user/item clusters where the assumption is that a user can be sampled from a user cluster and an item can be sampled from an item cluster. Locally you may have error incorrectly capturing the clusters but in practice this works fairly well and you dont have a one size fit all model.

If you're thinking about the cold start problem when you dont have any information about a user, yes it's possible that your overall statistics is a combination of many subpopulations that doesn't really fit anyone very accurately but there are ways around this as well.

So in short - a naive ML recommender that perhaps you get out of a textbook would certainly face these issues but the are addressable in production with an "industry" grade recommender system and will lead to better performance. I don't have firm numbers but in A/B/C tests compared to other black boxes I've seen 20-30% better performance when you start to address some of these issues.

>If you can't avoid say screaming in public that's a problem not just a non issue. Arguably the tolerances might be low, but it's hard to argue with the basic premise.

People are prescribed stuff for much much less than "screaming in public", including feeling stressed in BS high stress jobs, being "hyper-active" at school, etc -- with "modern psychology" more often than not siding with the normality of abnormal and abusive societal norms (as it always did. Science is conservative. Not so long ago you got medical treatment for being gay, for example).

>Granted doctors assume if your talking to them you have a problem. But that's arguably a reasonable prior.

Doctors also prescribe and over-prescribe all kinds of BS to people who don't need it, and get all kinds of pharma percs for doing so.

Heck, they'll go even to doing un-needed operations, compared to that, prescribing BS is nothing:

https://www.google.gr/search?q=uneeded+operations+scandal

That is way I said not constraint by your skeleton - or what ever else. The point is that at least some dimensions are obviously correlated, not that all dimensions have to be correlated.

Actually if anything this phenomenon would increase the probability of the super IQ thing. Intelligence isn't determined by one gene, but many thousands. Just like there are very few people who happen to be exactly average on dozens of different traits, there are very few people who happen to have the absolute best possible combination of genes to get max intelligence.

That is, assuming there is some optimal combination of genes, it gets exponentially more unlikely any individual has it, the more genes there are that affect intelligence. Just like you would be unlikely to flip 1,000 heads in a row. In fact flipping even 600 heads out of a thousand, is exceedingly unlikely.

While hindsight is 20:20, parts of this should have been more obvious.

> Daniels generously defined as someone whose measurements were within the middle 30 per cent of the range of values

> Daniels discovered that if you picked out just three of the ten dimensions of size ... less than 3.5 per cent of pilots would be average sized on all three dimensions.

30% raised to the third power is 2.7%. Basic probability. I guess everyone assumed there would be heavy clustering instead of largely independent variables?