I disagree with most proposed use cases here. What these models are good for is executing on well defined process where the process outcome can be evaluated to be correct.

E.g., if I ask AI to fix a bug, there is a way to evaluate whether the bug was fixed (I can reason through the code or run it to confirm).

When I ask AI to write a commit message, I can validate that the commit message describes the patch. I go through the code an check if it accurately labels the commit.

When I ask AI to generate an image of a tree, ..., you know the drill.

However, when I ask AI to do something much more open ended, e.g., "be the CEO of my project," "tweet for me," then it becomes rather problematic because for that the context window is e.g. 2 years and there's really no clear cut approach to evaluate on a per-task basis if the AI did a good job. It'll be very hard along the way to reflect or validate whether good decisions are being taken.

if it is a process where you strictly have to follow a protocol, and where the character of the human is explicitly not wanted in the outcome, then we should consider adding AI in the mix

Think about a government worker that strictly executes on process. We wouldn't want that worker to put their personality into the process necessarily, we'd want to execute based on criteria. That, in my opinion, are the tasks where AI can really help automate stuff.

Anything where your outlier-character trait helps you excel, maybe AI is suboptimal

Tim, I see what you are saying, especially as I'm re-reading this. I would probably rewrite the use-case area.

But, if humans are heavily involved in setting up the system, you can actually make each of these use-cases into well-defined processes, where the outcome can be evaluated to be correct.

To start off, you wouldn't ask an AI "be the CEO of my channel" you would ask it "analyze the channel activity of the last 3 months and suggest a list of 3 proposals to vote on." So, to start, it would be a time-saving help: instead of a bunch of individuals putting up hundreds of proposals everyone has to vote on, the AI would look at channel activity, based on likes and reputation, pre-evaluate what kinds of proposals the community would be likely to approve. Combine many of these kinds of time-saving tasks and you get "Governance genie"