doing things yourself avoids cargo-culting

What does it mean to 'be yourself'? Selves aren't a well defined thing. Ultimately everything about your self is shaped by your surroundings and environment. To always 'be yourself' without any responsiveness to social cues to be would be a bad and antisocial thing if you take it literally. But it's often even hard to know what your 'self' is. What would I do in situation X? I don't necessarily know that. Even sometimes as the actual scene is happening. I'm in a room with LongtimeCrush and I have to react in real time - what does my 'self' do? I don't know, and so I do nothing. (see also being yourself takes practice, which discusses a related point that even if you 'know' the high-level values that define yourself, it's not going to be automatic to translate those into action without a lot of practice).

One answer, from a normative perspective: a way of 'being yourself' that I can justify from higher principles. To 'be yourself' is to do things that you've found to work well for you, not things that you've seen other people do. and that's because watching other people is prone to cargo-culting: all the differences between you and the other people are effectively confounding factors. So you're doing causal inference from correlations alone, and will tend to miss the real causes. whereas with your own actions it's a controlled experiment. You try doing a thing, and not doing a thing, and this gets you an effect size, assuming all else is held equal. And because it's the same you in both cases, a lot more is held equal than when you move between your experience and someone else's life experiences. Things that you've found to work for you, in your life, are more reliable than what works for other people.

This is an argument for on-policy learning. If you observe other people's actions, then:

• They will be observed at a different $p(x)$ than yours---a different distribution of world states and perhaps actions. This is covariate shift. But also:
• They will have a different $p(y | x)$---a different distribution of outcomes given states and actions. This shouldn't be possible if our model of the world is really coherent. But almost always the real 'world state' is really x_visible + x_personal, where we may agree on some observed features x_visible but not on many other features x_personal.