The ideal training data for an LLM is not what you wrote. It's the full sequence of your internal thoughts and all the individual edits while you wrote it.
But you make do with what there is.

→ View original post on X — @karpathy,