Human Bias Stems from Overfitting

Humans tend to use their own complexes to explain some phenomenon, and to keep adding more causes one by one to elaborate on it.
And this kind of repeated learning often turns into one's own predictive ability (know-how).
The problem is that the more sophisticated that predictive ability becomes, the more its predictive performance drops.
This produces a phenomenon, much like in poker or the stock market, where beginners earn more money, and the more they (clumsily, while their experience is still half-baked) learn, the more they lose, or the more their predictions miss.

Just as with AI, when humans, too, accumulate knowledge (formal education, study) much faster than experience (learning or social life), they end up effective only in narrow (localized, specific) domains.

In such cases, when making predictions, instead of taking all of the variables into account and reflecting them, you should pick only the top one or two variables and use those as predictors.
It's a bit like the habit of, when buying a stock, simply deciding in advance that you'll sell it once it crosses some fixed percentage and then selling unconditionally when it does.


This idea is explained in more detail in Thinking, Fast and Slow.

Simple statistical rules outperform intuitive 'clinical judgment' - https://normalstory.tistory.com/m/entry/%EB%8B%A8%EC%88%9C%ED%95%9C-%ED%86%B5%EA%B3%84-%EA%B7%9C%EC%B9%99%EC%9D%B4-%EC%A7%81%EA%B4%80%EC%A0%81%EC%9D%B8-%EC%9E%84%EC%83%81%EC%A0%81-%ED%8C%90%EB%8B%A8-%EB%B3%B4%EB%8B%A4-%EB%9B%B0%EC%96%B4%EB%82%98%EB%8B%A4


Human bias, in the end, also stems from overfitting on experience or learning.


Experts (specialists) have the ability to capture details that ordinary people miss.
They consider parts that ordinary people fail to take into account.
But that strength becomes useless the moment they step outside a specific local domain.
So prediction becomes the work not of specialists but of generalists. (Of course, the generalists here aren't ignorant, ordinary-style generalists, but generalists who package together many specialties.)
Cases where collective intelligence cracks tough problems often resemble this kind of situation.


And maybe that is why
offline professional groups -
taxi companies couldn't build Tada or Kakao Taxi,
and bar associations couldn't build LawTalk.

By contrast, some online professional groups tend to have flexible cultures.
Google and Meta, while waiting for their own details and conditions to be perfectly satisfied, ended up building overfitted services or setting overfitted policies (business models). OpenAI, on the other hand, opened up first with a minimal model and so was able to become the first mover, while Google and Meta either become complacent or, embracing change, drag the runaway cow back out and turn the cow's stable into a cow station.
So it seems the companies doing well are the ones being lean-startup and agile.