The result was a bit worse than my result earlier and so I tried it in GPT-3.5 and then with a few other of their bots.

Seems like they are using user inputs to fine tune 3.5, which makes perfect sense and what I would probably build too.

I would be extremely interested for somebody to quantify the additional skill sets that are available in this corpus versus the broader corpus of whatever is going to be in the GPT4 and beyond data update

Sadly, my guess is that unless there’s some other draw beyond the individualized tone, then the massive general systems that get the most users will make these small efforts impossible to sustain.

A tangential thought does occur to me though:

If I was to ask my grandmother a question and she had a detailed answer for something that I know she had no experience in, I would question the validity.

So the fact that me asking one of these bots about something that is totally out of their wheelhouse, and they have a clear answer, is not something I would expect from a real person.