This is at least the third time in my life that we've seen a loudly-heralded purported the-end-of-programming technology. The previous two times both ended up being damp squibs that barely mention footnotes in the history of computing.
Why do we expect that LLMs are going to buck this trend? It's not for accuracy--the previous attempts, when demonstrating their proof-of-concepts, actually reliably worked, whereas with "modern LLMs", virtually every demonstration manages to include "well, okay, the output has a bug here."
I do seem to vaguely remember a time when there was a fair amount of noise proclaiming "visual programming is making dedicated programmers obsolete." I think the implication was that now everybody's boss could just make the software themselves or something.
LLM's as a product feel practically similar, because _even if_ they could write code that worked in large enough quantities to constitute any decently complex application, the person telling them what problem to solve has to understand the problem space since the LLM's can't reason.
Given that neither of those things are true, it's not much different from visual programming tools, practically speaking.
This is a feeling I have too. However, compared to Visual Programming it's perhaps harder to dismiss? Visual programming - its pretty obvious you can't effectively or even at all step through a UML diagram with a debugger to find a problem. The code that gets generated from diagrams, with visual programming, is obviously cr*p. So the illusion doesn't last long. Whereas AI - it kind of looks OK, you can indeed debug it, its not necessarily more complex or worse than the hideously over-complex systems human teams create. Especially if the human team is mismanaged e:g original devs left, some got burnt out and became unproductive, others had to cut corners and make tech debt to hit unrealistic deadlines, other bits get outsourced to another country where the devs themselves are great but perhaps there's a language barrier or simply geographical distance means requirements and domain understanding got lost somewhere in the mix. So, I suppose I sit on the fence, AI-generated code may be terrible but is it worse than what we were making anyway? ;). In the future there are probably going to be companies run by "unwise people" that generate massive amounts of their codebase then find themselves in a hole when no-one working at the company understands the code at all. (whereas in the past perhaps they could hire back a dev they laid off on large contractor rates of pay to save the day). Seems inevitable one day the news will be of some high profile company failure and/or tanking of stock caused by a company basically not knowing what it was doing due to AI generated code.
I use spreadsheets a lot, and often they end up serving as an MVP «prototype» until it’s obvious I need a proper database and custom UI, then build an application for the task instead.
Why do we expect that LLMs are going to buck this trend? It's not for accuracy--the previous attempts, when demonstrating their proof-of-concepts, actually reliably worked, whereas with "modern LLMs", virtually every demonstration manages to include "well, okay, the output has a bug here."