[atoav]

The problem I see is that LLMs have the tendency to pull everything towards the mean of whatever you ask it to. But the really, really good stuff is often totally out there on the fringes.

You can get some good results with really precise instructions, but this requires someone who has already thought a lot about what they are after.

[enos_feedler]

Coming up with precise instruction about what we are after isn't always possible. In many cases we only know what we are after once we see it. And if you can't express what you are after into tokens you can't outsource that choice to anyone else except your own direct experience of the process.