> as well as a typical human could I think this just isn't a good enough definition to be able to usefully reason about the concept. A typical human might say

> as well as a typical human could I think this just isn't a good enough definition to be able to usefully reason about the concept. A typical human might say "I don't know" or "I don't care" when asked a question, or say "I refuse" when given an instruction. Also, the typical human is a stranger, who knows nothing about you, and may have a completely different cultural background to you, so there might just not be enough shared understanding to be able to expect a meaningful exchange. Even if we adjust the definition to account for all of that, it seems like all relevant inputs would fall into one of three categories: simplistic questions/prompts that GPT-3 has already seen a hundred times before; complicated word/logic puzzles that many people would also fail at; or extremely technical requests like writing a novel which is a specialist skill that is best considered as a narrow domain to be solved by a dedicated AI. With that categorisation in mind, I think we can look more closely at what AIs are missing to reach the "verbal parity" level that the article cares about. Skipping the first category as too trivial, the second category hints that what an AI needs is a world model, and an ability to map statements from the text input into that model. For example, consider the logic puzzle of a father and his son in a car accident, and the surgeon saying "I can't operate on this boy because he's my son". A human might give up or give an invalid answer when asked to explain the supposed paradox, but if we require an AI to solve it for "verbal parity" then it will have to be able to build a world model in which there are accidents and injuries and surgeons and fathers and sons and abilities and limitations and emotions. My contention is that the ability to do that world modelling is what an AI needs in order to solve such puzzles in the general case, so if that's what we're saying verbal parity requires, then hidden within that world modelling requirement is basically AGI. For that not to lead immediately to AGI, there would have to be some line between world models that are only rich enough to answer simple puzzles, and world models that reach some critical level of complexity, perhaps enough for self-awareness or self-improvement, although maybe improving a world model is a very different skill, which might be where self-awareness risks gaining a sense of agency and goals. As for the third category, I think it is possible that an AI could learn to write books just from seeing how they are constructed from other examples, along with some careful programming devoted to maintaining causality and emotional impact throughout the narrative. Similarly there could be a set of rules for writing research papers or legal briefs from a given set of inputs, which might be no harder technically than playing StarCraft. Alternatively, once an AI has a sufficiently developed world model, the ability to imagine the story of a novel, or a scientist communicating their research, or a lawyer arguing their case, should fall out of the model naturally, and might just need the right prompt to generate it.