Problem with the theory is that people believe in LLM strongly enough that whatever pressure there is within a market to be vaguely similar evaporates. SQL certainly has dialects, but at least the basics are vaguely similar, as an example. Working with a vendor that is oddly different from every other vendor in the space and we applied pressure to implement more typical interfaces. Their answer was “just have an LLM translate for you and use our different and frankly much weirder interface”. When we did cave and use it and demonstrated the biggest LLMs failed, they said at least they give you the idea. Zero interest in consistent API with LLM as an excuse. On the write your code for you, it has to be kept on a short leash and can be nightly a nightmare if not overseen, though it can accelerate some chore work. But I just spent a lot of time last week trying to fix up someone’s vibe coded migration, because it looked right and it passed the test cases, but it was actually a gigantic failure. Another vibe coded thing took 3 minutes to run and it was supposed to be an interactive process. The vibe coded said that’s just how long it takes, if it could be faster, the AI would have done it and none of the AI suggestions are viable in the use case. So I spent a day reworking their code to do exactly the same thing, but do it in under a second. For the jira ticket scenario, I had already written a command line utility to take care of that for me. Same ease of use instead of using jira GUI and my works torturous workflows, but with a very predictable result. So LLM codegen a few lines at a time with competent human oversight, ok and useful, depending on context. But we have the similar downside as AI video/image/text creative content: People without something substantial to contribute flood the field with low quality slop, bugs and slow performance and the most painful stuff to try to fix since not even the person that had it generated understood it.