This... for my most important use case (applied numerical algorithms) it is in fact beyond not useful, it is negative value - even for highly available methods’

This... for my most important use case (applied numerical algorithms) it is in fact beyond not useful, it is negative value - even for highly available methods’ codes. Sure, I can ask for it to write (wrong) boilerplate but it is hardly where work ends. It is up to me to spend the time doing careful due diligence at each and every step. I could ask for it to patch each mistake but, again, it relies on a trained, skillful, many times formally educated domain expert on the other end puppeteering the generative copywriter. For the many cases where computer programming is similar to writing boilerplate, it could indeed be quite useful but I find the long tail of domain expertises will always be outside the reach of data-driven statistical learners.