Post #2741960

A pretty common explanation for bots being good at code is tied to the large open source corpus they (largely pirated) trained on. After a few weeks of using claude code to build a system that helps the llm/harness plan and maintain alignment to my intent, i can confidently say, that bitch fell on the head. I think LLMs are good code because code has structured intent. Predicting next words is def useful for using tools, and when you can lean on an AST, you can “”reason””. But try to get a bot to even format a coherent question much less understand a positive or negative control. Nope. #claude #llm #