> You cannot skim the AI's code. Every word of it must be examined.

I understand what you're saying. But be careful of this argument, it's susceptible to safety counter arguments.

"Well, this person thinks we shouldn't carefully review all code. Do we want him working on our super-critical FarmVille clone?" (That part is of course ironic. It changes if you're working on medical devices.)

A closely related argument is that the cost of the back-and-forth with a code submitter who submits buggy code is drastically higher. The hand-holding and teaching and encouragement is very expensive.

Maybe that's acceptable to the person who posted this article. But it's a possible avenue of resource exhaustion attack if you're not careful.

I'm sure AI will improve, with loving prompters. But some of the discussions around it seem dishonest, which is troubling (not you of course).