I haven't used claude, but I have that LLMs are great at generating code that is well defined with prior art but fail at anything a step beyond that. Building the UI for linkedin is simple. The hard part is managing the different roles and levels and state and micro-features that Linkedin has without hitting performance issues and bugs.

Right now, AI has the skill of a friend's son in high school that's good with computer has. I am really curious if these AI companies can overcome this level because I think hallucinations happen often when your requirements are too tight and the AI just makes stuff up because it can't solve the problem.