I'm glad you found the discussion helpful! To be clear, my concerns aren't about synthetic data, but about synthetic text, i.e. setting up systems where lawyer (etc) take the output of an LLM as if it were information.

→ View original post on X — @emilymbender