Explore has gotten way cheaper than exploit.

That happens when any new input changes.

It's less about LLMs being great at explore (though that's part of it).

It's mainly that LLMs are ushering in a new paradigm.