The fighter-jet seat problem is pretty famous.

https://www.thestar.com/news/insight/2016/01/16/when-us-air-...

The article you cited is an interesting just-so explanation but it's mostly narrative, not really correct. It's based on this book https://www.harpercollins.ca/9780062358363/the-end-of-averag... which espouses some rather non-quantitative concepts (it seems like the author had a hobbyhorse and they rode it as far as they could).

I don't know that it's "pretty famous", as I find only a few references to it (the book I mentioned, the article you mentioned, and a few other references).

I worked in biology for a long time and work with ML datasets that make money based on statistical analysis of people properties and the idea that "no one is average" is technically correct and also completely missing the point of scientific analysis (I definitely acknowledge that science doesn't do a good job with personalization).

> I don't know that it's "pretty famous"

I'm sorry you haven't heard of it but it's fairly well known.

Here's it on QI: https://www.youtube.com/watch?v=Ch9UuMFwzCE

It’s shown up on HN before at minimum.

My father was an instrumentation engineer for the DoD and much of what he worked on during his time at the high speed test track out at WSMR did involve ejection seat testing.

They had an assortment of very expensive mannequins rigged up with a multitude of sensors to determine strain or torsional effects that were used to determine whether the ejection was "survivable." I believe the certification process included the smallest possible dummy that could survive the seat ejection, and the largest that would fit.

Too small and the seat couldn't be rigged with enough ballast to prevent the acceleration of the ejection motor from crushing the pilot's spine. Too large and the seat wouldn't clear the tail or other aircraft structures. If I understand it correctly, the seats are configured to the pilot for this reason.

It's absolutely true that designing around averages is not how ejection seats are constructed or, for that matter, tested.

I strongly urge you to become familiar with the fighter-jet-seat problem if you're working with humans and applying averages.

"no-one is average" is more than technically correct. If your scientific analysis is missing this key point about humans, then it's basically worthless because it doesn't apply to real humans. This is what the US Air Force found: you have to personalise. You can never apply a generic result to an individual and have a good outcome.

I don't think he is misunderstanding anything.

Using the average sometimes works, some times doesn't. Number of arms an average person has is less than 2.

Definitely check out the fighter jet seat problem.

Mean is not a robust statistic, so that's why it's often better to use Median instead.

I did. Not impressed.

If 60% people perform better with light theme and 40% dark theme, average is going to be that light theme is superior. This doesn't make dark mode a bad idea.

This is going to appear nitpicky, but the average preference is some average gray, neither dark nor light. The mode is light. That relates back to the fighter jet thing: if you’re going to build a non-adjustable cockpit, build one that fits (within some tolerance) the most number of pilots, not the average on every dimension.

or, better, don't build a non-adjustable cockpit

IE, argmax, not mean

All you're saying is that means aren't informative if distributions are bimodal. Which is both totally true and completely missing the point.