When I am talking with Chat GPT or Claude about abstract concepts, I am often surprised by how they seem kind of dumb... like they aren't benefitting from their extra parameters over top open models like Kimi or GLM. In fact they often seem stupider than these open models that I run locally in the way they leap to conclusions and try to oversimplify concepts and produce more "not x but y" like slopisms My guess about why this is: system prompts are lobotomizing the model by trying to...