1 comments

  • kermatt an hour ago

    If people are not writing novel code, nor learning how to, does innovation slow since LLM generated code is largely customized copy-paste, and there's nothing new to train the models on?