1 comment

[ 0.21 ms ] story [ 20.2 ms ] thread
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?