> It cannot tell right from wrong It doesn't know which facts about the world are true and which are fabricated except through the text it's trained on, but to

> It cannot tell right from wrong It doesn't know which facts about the world are true and which are fabricated except through the text it's trained on, but to a large extent neither do you. It suffices to merely have the ability to reason about it. The primary difference is that whereas you reason fairly competently from one perspective with one pseudo-coherent set of goals, GPT-3 reasons to a weak degree from all perspectives, privileging no view point in particular. How important this ends up being depends on where the model plateaus. On one extreme, if it plateaus close to where GPT-3 already is, no harm done, it's a fun toy. On the other, if it scales until perplexity gets to far-superhuman levels, it doesn't matter at all, since you can just prompt it with Terence Tao talking about his latest discovery. Naturally, it will land somewhere in the middle of these points. The question is ultimately then whether the landing point captures enough general reasoning that you can use it to bootstrap some more advanced reasoning agent. A sufficiently powerful GPT-N should, for example, be able to deliberate over its own generated ideas and sort the coherent reasoning from the incoherent.