LLMs fill in underspecified parts of the users' request with frog DNA.

The frog DNA is inherently mushy; average.

This means that the under-specified parts become more average, pulling toward mediocrity.

That's bad… but also, without that frog DNA filling in the gaps, the idea might not have been fully specified enough to be viable.

It's better to have a viable request that you can interact with the results and tweak and specify better, than to get a brick wall of "sorry, does not compute".