Honestly, that’s a smart thing for AI companies to do. AI is surprisingly decent at extrapolating from existing codebases, but it’s useless at starting from scratch. If one model says “I can’t do that, Dave” and another spits out garbage, you’re getting the same amount of useful code out of both and a much better signal-to-noise ratio from the first.
imo it’s the opposite, AI is good at starting projects by giving you boilerplate code, but bad at considering the full context of an existing project. Better to be doing the larger structure stuff yourself and only giving the LLM self contained tasks.
Honestly, that’s a smart thing for AI companies to do. AI is surprisingly decent at extrapolating from existing codebases, but it’s useless at starting from scratch. If one model says “I can’t do that, Dave” and another spits out garbage, you’re getting the same amount of useful code out of both and a much better signal-to-noise ratio from the first.
imo it’s the opposite, AI is good at starting projects by giving you boilerplate code, but bad at considering the full context of an existing project. Better to be doing the larger structure stuff yourself and only giving the LLM self contained tasks.