Try making a webpage using an LLM when you don't explicitly tell it to optimize for page weight. Not that developers get to spend much time optimizing, but they can actually think and make reasonable choices of how to build it and will notice if it's noticeably slower in e.g. the integration environment compared to localhost (even if I'm sure they have good machines and connections compared to 90% of users, at >10MB you start to notice, whereas token predictors do not have a concept of time while fetching the page in a test run)