For me the models are great creative thinkers, and good at analysis of text, but doing anything that must be accurate such as code or maths or physics etc, I just don't trust them.

I would not trust a model to produce a block of code that I then only test before deployment. It feels too much like when one outsources coding and you get back a tested unit, but under the hood it is delicate spaghetti code.

I'm sure the models will get very good at the thinky part one day, and when they do there'll be a vertical line in development of everything, but until they can do the thinky part I just can't bring myself to spend my money on them because I just don't trust they'll be correct 100% of the time.

That being said, us humes are also not correct a 100% of the time. Code without bugs is a very rare thing, so perhaps I'm being too hard on the models and the issue is entirely my bigotry against silicon intelligence (that I didn't know I had till this reply)