I love some of the ideas you pointed out. > Precise but not accurate. Since it is always open to interpretation, it simply cannot be used for verification Agr

I love some of the ideas you pointed out. > Precise but not accurate. Since it is always open to interpretation, it simply cannot be used for verification Agreed - the main way I try to manage this is to have a very coarse interpretation of statements - almost like a probability distribution of semantic meaning[1]. For example, if someone says "I like spaghetti", most of the probability distribution is located somewhere between "I tolerate spaghetti, but I won't say it that way in case you're about to serve me spaghetti" to "My favorite food is spaghetti, but only if it's served in the general style of my hometown". I might be misinterpreting what you're saying right now, but we can continue to have this conversation, and I hope you're enjoying it even as I am enjoying it! I generally agree with the idea that it cannot be used for verification, depending on what you precisely mean (my probably distribution of _your meaning_ has significant overlap with what I would agree with, but not entirely, though to be pedantic, it never entirely overlaps since there's always a chance of misinterpretation). I think that's why it's important to establish relationships and baselines of communications. Repeatedly successfully interpreting each other's words can provide a strong baseline for future communicative success. It's important not to project one relationship onto another, which we are _very_ prone to do. E.g. we should not project how we communicate with one sex to another sex, or necessarily even one person to another person! [1] Maybe this is the next representation of semantic meaning that LLMs will use somewhere in its internals? Instead of a single embedding, it describes some probability distribution around an embedding with the greatest probability?