Mmm the most impressive thing I see LLMs do is take a piece of unstructured input and transform it in some way - summarize, extract information as JSON, etc. Th

Mmm the most impressive thing I see LLMs do is take a piece of unstructured input and transform it in some way - summarize, extract information as JSON, etc. This wouldn't be possible if it were repeating training data verbatim, since it works on new novel inputs.