The biggest problem is LLMs’ sycophancy and initiative - Claude code assumes that it has access to certain tools that I have (intentionally) not given it access to, and it won’t steer people towards the right answer, unlike search engines, which may have been wrong but their information was correct