I wouldn't be surprised if #GPT-like personal assistants became as commonplace as smartphones.
They're already more useful than Google for questions.
Can help with writing by brainstorming and filling in fluff.
They can be a command-line interface for non-nerds. Personal assistant for Excel, db queries, file management tasks, software configuration (except printers, they're possessed and can't be helped.)
My impression is that all of these models are useless at generating true statements. E.g. I ask chatgpt "What is the halflife of aspirin in blood?" and it tells me it's 20minutes (actually it's between ~2h and ~10h depending on dose). There's no way to tell, based on the answer alone, that it's wrong: it's not outrageously wrong (well, given the dosing recommendations it does sound low, but not impossibly low) and the answer is grammatically correct. So, to verify the answer I actually need to go to as much effort as finding it would take.
@robryk Yes, wrong information is problematic.
However, the accuracy could improve. Earlier GPT would answer any absurd question. ChatGPT is already able to spot nonsense questions. I think it's a matter of more data and more development to let it be more accurate and more aware of its limitations.
Actually, a better example (granted, it's harder to find them but it feels like there's some knowledge-graph-style preprocessing happening):
> Which is larger: dog or a year?
> A year is much larger than a dog. A year is a measure of time, while a dog is a living animal. A year is typically 365 days long, while the size of a dog can vary greatly depending on the breed. Some breeds of dog can be quite small, while others can be quite large, but even the largest breeds of dog are still much smaller than a year.