I'm not sure if that will hold up. The LLM is not going to do anything random and that is actually a powerful component that makes original output possible.

▲

kyralis a day ago | parent [-]

I wonder if a combination would be useful. Use an actual GA to do the mutation, and then let an LLM "fix" each mutated child.

	▲	jacquesm 19 hours ago \| parent [-]
		Could be. But the interesting thing is that all you can do here is optimize. Random chance is - like attention ;) - all you need.