I’d frame it a bit differently. LLMs are pretty good at generating the midwit solution to problems because that is the bulk of the available training corpus. It

I’d frame it a bit differently. LLMs are pretty good at generating the midwit solution to problems because that is the bulk of the available training corpus. It is a generic “best practices” generator. You would expect it to be better than a third of human architects almost by definition. On the other hand, they are pretty poor at reasoning from first principles to solve problems that are far outside their training corpus. In some domains, like performance-sensitive platforms, the midwit solution is usually the wrong one and you need highly skilled people to do the design work using context and knowledge that isn’t always available to LLMs. You could probably use an LLM to design a database kernel but it will be a relatively naive one because the training data isn’t available to do anything close to the state-of-the-art.