> The problem is, a computer has no idea what "improve" means unless a human explains it for every type of problem

Not currently.

I don’t really think AGI is coming anytime soon, but that doesn’t seem like a real reason.

If we ever found a way to formalize what intelligence _is_ we could probably write a program emulating it.

We just don’t even have a good understanding of what being intelligent even means.

> The best we'll get, I believe, is an AGI that can optimize itself within a few narrow problem domains

By definition, that isn’t AGI.