I repeatedly feel like the same stuff is being missed, on this topic.

Mostly, what a skilled human brings to the table, is 1. breadth of 2. non-direct-business problem domains.

When you ask me to build a Widget, you express domain-specific constraints and problems to me. You need to express those to an AI, too. In both cases (ideally, at least; modulo “they’re still in the oven and have a ways to go”) you, the business-domain person, gets out A Thing That Does What You Asked.

I, though, know about all the things you didn’t know to ask. I can spot, in your description or during implementation, ‘that sounds suspiciously like distributed state.’ ‘That’s going to require accessibility review.’ ‘There’s going to be a hot-loop here requiring especially sensitive performance-regression testing.’

There’s an array of, I dunno, prolly around 15-20 “things I know about,” after years of doing this. And, while it seems LLMs can speak effectively on any one of those, I constantly find them breaking down on the reasoning/analysis meta-task of *knowing what the delicate balance between those things needs to be, at this particular point in the architecture.

Can I invoke some Super-LLM, and spend two hours enumerating every design-dimension I want to be handled carefully? Yes.

Can I apply human direction, and tell an AI to review from one of those particular angles? Also yes.

But getting an AI to keep all of them in its head at once, while architecting, and efficiently deploying the limited resource of reasoning/attention/time/testing/LoC-needing-maintenance-for-a-decade into an efficient allocation between those domains, has continuously seemed impossible for me. (And I have been putting a ton of effort into this, because I fucking hate feeling left-behind.)

Here’s the thing: I, mostly, get the feeling that other experienced engineers are on a similar wavelength (both pro- and anti-camps. And, not PMs/C-suite types; they’re often Full Koolaid in a silly way I’m not addressing here.)

But where the two camps differ most (… to come around to finally answering the OP?)

… is whether one believes you can solve the shortcomings of one LLM, by applying another, wholesale.

While I feel I’ve mostly seen both camps agree with me that, “yeah, sure, when I ask these Amazing LLMs to build me a thing, it’ll often have, say, security holes” … I find that the pro-AI camp tends to come back with “… but that’s the amazing part! I can just tell another LLM to rewrite it, and focus on security this time, and omg it totally does!!”

Which is, to me, just so totally missing the point. Because there’s limited attention-budget, whether you’re a human being or a GPU with a context-window; and re-engineering the entire thing from scratch means you’re going to completely detune the other parameters.

Put concretely: yes, we’re getting close to the point where you can “just tell the LLM to make my app accessible.” But we aren’t, and I am mildly convinced will-never-be, to a point where you can “just tell an LLM to make an app accessible, and oh by the way don’t forget-about-and-fuck-up the sequence-calculus correctness under CAP, don’t introduce any new security holes, ensure the parts that need to be performant stay performant, and thread observability correctly through all components without leaking PII, while you’re at it.”

And you totally can tell a human to do that last part. That’s literally what a human is. An intuitive-leap-making-machine. “This function I’m fixing for an unrelated ticket, shit, this will leak PII.”

We’re not perfect, either … but so far, my personal experience still leads me to feel there’s a qualitative, not quantitative, difference to how human engineers (or even beginners!) I’ve worked with, versus my experience watching Claude 4.7 1M Pro Ultra Max Reasoning Genius Mode unspool its stream-of-slop consciousness.

tl;dr the way LLMs need babysitting doesn’t lend me to believe they can reasonable be a net productivity bonus, even if they get significantly smarter/more-correct in narrow, targeted goals; because they seem to do an inherently poor job of intuiting correlated goals you don’t (know to) enumerate for it; or even of balancing multiple goals that are inherently in-tension with the user’s stated goal. (“you can’t do this securely, you can’t do this and maintain accessibility, you can’t do this under CAP …”)