From what I understand, LLMs can't really generate anything meaningful that doesn't implicitly rely on the operator's choices. It's hard to make the right novel

From what I understand, LLMs can't really generate anything meaningful that doesn't implicitly rely on the operator's choices. It's hard to make the right novel choices as soon as you leave well-defined problem spaces. In terms of math and biochemistry the cost of generating candidates has collapsed, but the cost of validating them hasn't.