Just some things that GPT3 can't do:
- A large portion (of at least mine) queries are 'grep the internet' type of queries. GPT3's powers lie in directly opposite part of the spectrum.
- GPT3's output can not be trusted. It is basically a statistical output of the most likely words to be responding the prompt. In many cases it will be complete gibberish and you have no way of knowing (programmatically) that it is so that you could potentially replace it with something else.
- GPT3 is notoriously bad with numbers and dates.
Finally Google has managed to redefine what search means to the point that 'to search something' means 'to google something' and even if you had superior (by what measure?) results, if they don't match what we are used to getting from Google, you might have a barrier to adoption.
So my bet is that NLP models will augment the abilities of search engines, not replace them. We see that happening already.
Has it? From what I've seen it's just an autocomplete that looks natural.