People’s inner simulator is almost always more accurate than their explicit models. It’s just less precise. The thing about your statement of [if it were more accurate, people would be using it more, and be successful at more things] requires a few initial assumptions to be true. The first is that people are able to use their inner simulator on purpose, which is usually not the case from my observation. The second is that people are able and willing to take the path indicated by their inner simulator when it contrasts with their explicit models, which is also usually not the case. Then there is an additional factor where the inner simulator can (and often does) output a result that is contrary to the facade a person is trying to keep up for social reasons, which provides an additional impetus to reject the inner simulator (even if it is indeed more accurate, and dropping the facade would be the best way to produce the desired long-term result).
In regards to learning/picking up soft skills. People do indeed pick up soft skills automatically. Soft skills are notoriously difficult to teach. Most people that learn a soft skill learn it on their own, through a long-term automatic process, after taking in information that sparks a complicated processing of that information. I’ve never met someone who learned a soft skill without it being at least partially automatic. There is additional complexity in this in that people are “born” with an “affinity” for certain types of soft skills, and have an extreme amount of trouble learning any soft skills outside of their affinity range.
The actual reason learning soft skills involves reading some amount of “wrong ideas/facts/information” is because the processing of information for soft skill use is different from the processing of information for explicit model use. Or more accurately, the mind has multiple ways of processing information, and is not at all limited to a neuron-only model. The method of processing information for soft skill use is more of a resonance/antenna model, which benefits from additional points of information no matter how wrong they are (as long as they are at least distantly related to something that has at least an iota of truth), up to a limit of what the resonance chamber can hold coherently.
Last, your two points in that thought experiment are explicit model only, and do not relate to soft skill learning or use. The difference would be that you have an explicit use in mind for each experiment. A soft skill necessarily has a soft use. The difficulty in translation of “flattering people doesn’t help” is that such things are actually a general-scope statement, and they have to be for soft skill learning and use. So your thought experiment is akin to: “You read that white men can’t jump. You ask a white man to jump, and he is successful.” You’re combining a general-scope statement with an explicit-scope experiment.
Soft skills are hard. I’m extremely good at learning soft skills. I try as hard as I can to teach one whenever someone desires one, but they’re damn near impossible to teach. I’ve thus far only been successful at teaching someone how to do something related, in a way that their mind can grasp and run with, so that a few years later they will have developed the soft skill “on their own” (At which point, they will be doing exactly as I said exactly how I told them to, and they will tell me about this cool new thing they figured out how to do, and question why I didn’t ever tell them to do it that way).
So rather than teach how to learn soft skills, I’m going to describe a few of the most useful soft skills I have.
Maximize information input in a way that maximizes timeless usability. It works like this: prioritize observations in opposite order. The observation with the highest priority is the observation that is most distant from the desired observation while still being distantly related. The observation with the lowest priority is the observation that was first intended. For example: when performing a scientific experiment, the lowest-priority observation is the result of the experiment. Literally everything you could possibly observe related to the experiment is of higher priority than the direct result of the experiment. The reason for this is that distant observations have a tendency to be more useful for the future than direct observations.
Minimize friction to maximize usability at detail. When welding, move the welding rod and heating device at the exact rate that keeps the rod molten, and the welding pieces not-molten. When using a common carbon fiber dremel cutting blade on metal, move the blade through the metal at the exact rate that just barely melts the metal in the direction of the cut. When hammering a nail, make the nail move at a consistent rate that is fast enough to slide past the wood, but slow enough to not split the wood (this is like air resistance at sub-sonic vs super-sonic speeds). When talking to people (assuming you want to maximize usability of the other person’s mind), make statements and ask questions at a rate that doesn’t overload the person’s mind. When drawing a picture with a pencil, move the pencil across the medium at a rate that glides across the medium without catching on the medium.
Inference jump to maximize efficiency of skill gain. When developing a theory from a conceptual base, assume a related postulate to be true so that you can obtain more data from experiments. The postulate is most likely not true, but that’s not important. Experimenting under the assumption that a postulate is true is very useful in determining why said postulate is not true. Believe nothing. Treat all beliefs and assumed truths as inference jumps, because doing so maximizes their use. In order to make inference jumps with highest usability, all previously beliefs should be considered irrelevant. Failing that, consider the possibility that what you know to be wrong, or believe to be wrong, is in fact true, and use that as an inference jump.
I personally think the conceptual base of Less Wrong is contrary to efficient soft skill development. However, I think the conceptual base of Less Wrong is a potentially good platform to use to begin soft skill development. The most efficient way I can think to do so is to learn to use the conceptual base of Less Wrong, then make the inference jump that everything you’ve learned from Less Wrong is uselessly inefficient. This is the pattern I see everywhere people strictly adhere to a conceptual base. It’s almost always a good platform to expand further from, provided the expansion stops using the platform as soon as possible.