and yes, building a good harness that has good intentions is probably still the biggest challenge
claude code still unbeaten, it is too much of a good harness
me: please translate this document
gpt 5.6 sol: let me start by downloading a 6 GB local llm to do this translation
me: wtf, why don't you translate it yourself? is the local llm better at translation?
gpt 5.6 sol: no, best would be a frontier model like gpt 5.6 sol. do you want me to send this document to the openai API?