I may be an unusual audience but something I've appreciated about these models is their ability to create unusual synthesis from seemingly unrelated sources. It's like if a scientist read up on many unrelated fields, got super high and started thinking of the connections between these fields.

Much of what they would produce might just be hallucinations, but they are sort of hallucinations informed by something that's possible. At least in my case, I would much rather then parse through that and throw out the bullshit, but keep the gems.

Obviously that's a very different use case than asking this thing the score of yesterday's football game.