@mos_8502 I think, or maybe merely hope, that having deep expertise and actually understanding all the code being spewed forth at alarming rates (even if I haven't actually read most of it), has long-term value.

But, I don't think I have any choice about using it and staying employed. If I want to keep working in the field I'm in, I'm going to be doing it with AI assistance. That was up in the air until recently...but, the current generation models+agents lay the question to rest.

@swelljoe My reply to the vibe coder in this scenario is "well, then, what do we need you for? Pack up your desk."

Now, this is not to say I'm 100% dead set against it per se. I would restrict its use to expert-level programmers who understand the problem domain, on the grounds that you need to know the language at a high level and understand the problem if you're going to debug and maintain the code.

@mos_8502 the vibe-coder answer to that is, "I don't need to understand it, I just ask Claude to maintain it, too." And...I scoffed at that idea in the past. But, it's gotten really good.

I think it's still very dangerous in the hands of non-technical folks. And, I think most of the models are still flailing a lot (the frontier models are all pretty good, but Opus 4.5 is the one that Just Works, but all the lower end models are stupid a lot).

I dunno. The ground shifted beneath my feet.

@mos_8502 I was dragged into using it at work, and I've had to accept that I was wrong about how well it works. I still have plenty of reservations about its use. But...it can write good software very, very, fast. And, in an agentic configuration (like Claude Code or whatever other agentic harnesses you wanna use), it can test its own work, verify outputs, etc. It's genuinely astonishing how well it works.

Whether I like it kinda doesn't matter. It works well enough to where it will be used.

@swelljoe What worries me is the "competent operator". Best case scenario, the model generates perfect code that does exactly what the fuck it should -- and the user who prompted it has no clue how to maintain or debug the code.

@mos_8502 I wasn't enthusiastic about AI for coding, but Claude Code with Opus 4.5 is remarkable. I can't really say I'm happy about it, but given a competent operator, it produces quite good, if a little verbose, code that works the first time most of the time. It writes better and more tests than I've found most human coders do, it often includes accessibility markup without being asked (another thing human coders simply don't do), and it documents its work.

@nickzoic @mos_8502 A million lines of actual code, regardless of where it came from, is a catastrophic failure.

@mos_8502 It's like trying to accept code from a contributor who keeps making a gratuitous wrong change every time they fix the last thing you told them to fix, and trying to slip it past you. Infuriating and pointless when you could just do it faster yourself or reject their proposal entirely.