Same, exact experience here.

This one non-technical PM guy at work used Codex to develop a project I was expecting would fall on my plate. He asked me to do a code review on it. What it produced was riddled with SQL injection vulns and the UI was complete garbage.

Off of that example, the key stakeholders on my project are demanding I start vibe-coding everything. I raised the security flag and now they are saying, "well, now we have a prototype and real development can continue," but it's clearly just to mollify me and make me shut up, because no such development effort on that other project has been planned, scheduled, budgeted, etc. They are kind of just sitting around on it, hoping they can get everyone distracted long enough to sneak it out the way it is.

"But he did it in a week!" Yeah, it would have taken me only a week to make whatever of value actually was in that project. The reason our software projects at our company take longer than a week is not because of code, it's because we have an IT department that blocks production deployment of everything unless you literally get the president of the company to make them do it. That's not a repeatable process that every project can leverage.

There was another project another more-technical-but-not-a-developer guy (he knows how to use MS Access) did in Claude Code where, yes, Claude could read a bunch of PDFs he got from the client, get the salient details out, made an Access database out of it, and made a static HTML website out of it to make those documents easier to search and navigate. But again, the UI was complete, unadulterated garbage. And, the best part, he spent several weeks on just getting Claude to reliably process the entire set of documents. He never could quite get it to end-to-end do the entire process. It kept missing documents and reprocessing the same ones over and over again. A for-loop to iterate over a directory of files would have taken 2 minutes to code by hand and he got stuck on it for over a month.

AI will speed us up, my ass.

Look, if AI means I never have to open another PowerPoint from a client to read a "quad chart" on one particular slide to get the data I need to do my project because my client doesn't understand that PowerPoint is not a data transmission format, fine. I'll be happy with just that: AI vision as a library I can call out to from my code, just like we've been trying to do with OCR but traditional OCR sucks at the job. But there's a bigger drumbeat than that and it ends in dilettantism and laying off the junior analyst and developer staff. I will be no party to that.

I'm not even a front end guy, and have little experience with UI/UX, but its wild how easily decision makers are impressed with UI spit out by an LLM. This era of anybody being able to make a dashboard with Claude Code has made me really appreciate the amount of sweat that designers and devs put into a good user experience.

I agree with it being great for OCR, the most impact LLMs have had for me are structured outputs I can call from a function: "extract X value from this ambiguously structured document and return json that can my code can deserialize into Y type." However, how many people are doing something similar and spinning up $500k in GPUs just to avoid writing regex.