All that's missing is a reranking step that rewards incisive and non-obvious answers. Actually, some models are already implementing such a thing: LAmDA will prioritize contentful answers above statistical probability, the same way humans do.

GPT was only trained on "text" with no specific distribution, and specifically to predict a masked word from its surrounding context. As a result, the only questions you can use GPT to answer are of the form "sample from the most likely continuations of this text". It turns out a lot of problems can actually be posed in this form, if you understand the input distribution well. But the future probably looks less like models that operate directly on a distribution of language itself, and more like models that use their language knowledge to predict the volition of the person who gave it input, relate the input to their knowledge of the world (learned from corpus), and then use their language knowledge to convert an internal abstract representation of logical reasoning to something the user can read and understand.

I don't think the tech is extremely far off, it's probably a natural continuation of the current research.