Some initial notes/thoughts on GPT…
A useful (essential?) intro to what prompt engineering actually is:
Prompt Engineering vs. Blind Prompting
(Spoiler alert… blindly smushing data/tasks into chat.openai.com is not it)
Shared by Alex B…
Yes but.. Can ChatGPT Identify Entities in Historical Documents?
👆 This is one of many “hot takes” appearing on Arxiv currently (April 2023). I think the authors haven’t fully understood how best to use GPT for their purposes however. (See article above). In particular there appears to be an assumption that GPT should have an implicit “world view”, which it doesn’t. It’s therefore entirely unsurprising that GPT underperforms a specialised model in zero-shot mode and given no useful context.
A couple more “hot takes” of interest:
A Categorical Archive of ChatGPT Failures
A Multitask, Multilingual, Multimodal Evaluation of ChatGPT on...
… and this one has been causing a bit of a stir – not unsurprisingly given its frothy title 🙂
Sparks of Artificial General Intelligence: Early experiments with GPT-4
(It also presents a pretty good intro into GPT-4’s strengths/weaknesses, and abilities to reflect/correct).
LLM-related, this might be worth investigating for those with an interest in ethics, bias, etc. in relation to LLMs: