This aligns very well with a model I personally use. When I'm unsure, I complete the sentence in my head "I should do _thing_ but I want to do _other_thing_"

This aligns very well with a model I personally use. When I'm unsure, I complete the sentence in my head "I should do _thing_ but I want to do _other_thing_" If I'm still unsure, I try swapping the blanks in the sentence around a few times and I see which one feels more true, which one resonates more. That normally helps me see which thing was something that I felt like I should do, and which thing I actually want to do. The separate questions here are how to learn to access more of your inner intuition about what you want, and I think it mostly comes from practice of seeing if you can notice and using the method I mentioned above, and then seeing how you felt about the result. There's also a separate question about how to make sure your intuition is well informed, which I think comes from a combination of: trying a lot of things in life so you have many data points, being well rested and generally treating your body well (good food, some exercise), and being mentally healthy (calm, composed, peaceful). Our sense of what feels right is kind of like what our subconscious mind thinks would be best for us. And our subconscious mind is exceptionally powerful, it's fantastic at making predictions about short term emotional outcomes, it's seen everything that's ever happened to us and how it turned out, and so on. It'll serve you well to have a better relationship to your subconscious and vice versa! Some side notes that I could expand on: * "You" are not you, I think we are mostly our subconscious but our thoughts and words are the tool it uses to communicate and interact with the world, and to plan * Subconscious mind : conscious mind :: Pre-training : RLHF * The conscious mind, the part of us that has thoughts, is more like a limb, or a sense. It's more of a tool to interact with the world, and to receive input from the world, and the actual 'us' is something deeper, it's the part that has the 'wants' * RLHF messes with the ability for GPT-4 to be well calibrated. I suspect that's because of the "should" vs "want" / perception of other people's perceptions introduced during RLHF. * I wonder if psychedelics help access the subconscious mind / "base model without RLHF" and I wonder if this is why certain similar patterns seem to be noticed across different people who do psychedelics - because they're accessing the things that their subconscious mind has concluded, and many people's subconscious minds reach similar conclusions because we're using similar data. * A person who reads a ton of books is kind of like doing a "pre-training" phase. * Aligning a language model can borrow lots of methods from effective parenting (by effective, I mean where both the child and the parent end up happy), clicker training of animals, and effective delegation and management of employees. * Good advice is advice that helps people "want something they feel like they should do but don't currently want to do" or something that helps people "feel like they should do something that they already want to do but feel like they shouldn't do" * If you follow your "shoulds" you end up living surrounded by things you felt like you should do * Giving advice to other people is often your subconscious mind telling you what it wishes you would do (and so whatever you say to the other person is really either something you feel like you should do, or something you feel like you want to do)