But the numbers seem credible and realistic when you think about OS usage at that time. If it was hallucinating from web-pages that only go back to 2009, how on earth did it come up with such credible figures?
It's not looking things up, it is making things up. This is what it was designed to do... that's why they call it "generative AI"
Half of statistics are made up, so the saying goes. It's easy to make "credible and realistic" numbers if you understand confirmation bias. You want to believe in ChatGPT, so you see its numbers as credible and realistic, despite them not being so.
Also understand that we are in the beginnings of a hype cycle. There was a time when HN was exuberant about the potential for blockchain. I expect that our expectations for LLMs to be made realistic with time. I also believe that LLMs will provide value, unlike blockchain, so the hype cycle is probably closer to dotcom in how it plays out.
The plugin system will be interesting to see play out. It's sort of a hybrid system, using LLMs as an interface for intent while calling out to other systems for accurate retrieval. I'm curious how far it can be pushed, it's hard to see these things being able to interact and navigate all of the tools we have created. Too much variability and opaqueness.
It's not really. I don't know the exact figures, but it seems like common sense to think WinXP would have very low usage in Jan 2022, given it was released in Oct 2021. And anecdotal observations after living through a time and using computers a lot seems like a reasonable basis to gauge the rough correctness of stats.
not sure I am following you any more, but pretty sure XP is way older than Oct 2021
Your experiences were biased by being someone technical, windows was widespread, I know of a place still running XP as recently as Oct 2021 tho... there are some business that don't take updating seriously, so what you saw is not evidence of the general population.
Back to the point, it's best not to rely on your feelings when it comes to judging ChatGPT's accuracy