ChatGPT’s word statistics learns logic so well, for novel input it’s almost as good as intelligence
Over the last few months I’ve been defending ChatGPT & GPT-3 as a useable major breakthrough in NLU (natural language understanding). Coming as a fan and practitioner of both large language models (LLMs) and symbolic AI.
I do get the GPT skepticism.
But here today I try from one more angle to point out that GPT word statistics is so decently robust at logic — especially if you keep to a single paragraph prompt — that except for truly mission critical applications it just usually works.
Whatever the case, it’s an incredible end-user tool for research of all sorts and a harness-able NLU technology for use inside apps.
With care, intelligent assessments and decisions can be extracted from GPT.
So for anything that can tolerate the odd muck up — e.g. end-user research — GPT is the way to go.
Only it’s gonna keep getting better every six months.
Quick re-cap
I’ve approached the GPT debate from up to 10 point of views (LOL) to try and present a balanced view of LLMs that, despite their limitations and the skeptics, the merits far outshine anything symbolic AI has produced and that understanding and…
