Comment by abustamam
6 hours ago
> Then why wouldn't Claude do what it does every other time and pull from it's corpus of knowledge?
I dont know why, but I do know know that it doesn't. I know because I've tried similar things when playing around with AI to see what it could do. "Make an app similar to X app [link to repo]"
Depending on how its feeling, it may just read the readme or it may go through the codebase. But I've never gotten line by line code copying.
> I would expect the human-in-the-chain to verify what is going on
I agree with this, to the extent of your second point
> they are responsible for any output they lay claim to.
How far does the rabbit hole go for verifying what is going on? If the LLM did indeed copy an entire codebase at its own behest, what reason would the author have to verify that it wasn't copied, given that every other time they've used an LLM to code it never copied an entire codebase?
A bit outdated, but IME never copies things verbatim because of limitations in context. Even when you want it to be a verbatim copy, like in a migration, it struggles a bit to remember everything exactly from memory and always has to cross-reference.
Might be better in newer models though.
But for a real "copy" it would default to calling `cp` in bash.