Is it though; what does the term mean; are we talking to our computers, not reading anything and just judging the non-coder endresult? Or are you just using it to help coding. Vibe coding would be the former and while it sometimes works, it doesn't in many cases, including LoB tools if the backoffice is large and complex enough. Unless you have some interesting tooling to help out, in which case it is possible, but that's not really vibe coding at the moment.

I now have 5 "vibe coded" programs in our production chain at my non-tech company. Some have been running for over two years now.

I think what tech companies miss is that customers buy these expensive feature packed software suites and then use 3% of the functionality. Cutting edge LLM's, and especially the newer thinking ones, are incredibly talented at creating programs that offer 3% the functionality of software suites.

Just a few weeks ago we skipped out on a $1k/mo CAD package because claude wrote a program with a simple GUI that enabled our engineers to do the one task we needed the custom CAD software for (Manually creating part placement files from 90's era manufacturing files).

I think throwaway programs are fine here; you make them, ship them and that's it. But i'm quite sure you didn't vibe code them; you used an LLM to write a lot of code but helped it along; vibe coders cannot do that because they don't actually understand what the LLM did. It just worked, or didn't work, so they copy back and forth errors (if there are any) until it works. Using LLMs to assist writing software, sure, we do that every day and it ships. But if we don't help the LLMs when they are stuck, they don't get unstuck. They cannot solve some surprisingly trivial things which vibe coders will be stuck with.

> are we talking to our computers, not reading anything and just judging the non-coder endresult?

more or less this, yes. a quick glance, maybe skimming through a tricky part. otherwise, if it works, it ships.

>Is it though; what does the term mean; are we talking to our computers, not reading anything and just judging the non-coder endresult? Or are you just using it to help coding. Vibe coding would be the former and while it sometimes works, it doesn't in many cases, including LoB tools if the backoffice is large and complex enough. Unless you have some interesting tooling to help out, in which case it is possible, but that's not really vibe coding at the moment.

I spent 10 years of my life as an FE with a total obsession for the field, and took huge pride in my hand tuned HTML/CSS interfaces and well architected JS. I was "the UI guy" that anyone would come to with questions, as it was my expertise and what I loved to do all day.

Now I'll never write a line of that stuff ever again as long as I live.

I do not write code anymore. Everything goes through an agent. Occasionally I will flip a config value here or there, or clean up some leftover logs or whatever, but I don't type code at all anymore. My job is now to guide the agent and review its' PRs. Yet, I now easily have the productivity of an entire team from a few years ago.

You can keep your head in the sand over this stuff for a little while longer maybe, but you will be completely obsolete if you don't adapt soon.

Yeah not every developer does html templates for wordpress.

Of course if your job is extremely simple and generic, AI can kinda figure it out.

>"Yeah not every developer does html templates for wordpress. Of course if your job is extremely simple and generic, AI can kinda figure it out."

Keep telling yourself this. It's only hurting you.

I think anyone who isn't working heavily with agentic systems right now has absolutely no clue about the revolutionary things that have happened in the last six months or so. Our industry is going to be in a completely different universe this time next year.

"Past performance does not guarantee future results"

You can't extrapolate from the last 6 months, especially with technologies this new. There's no Moore's Law of LLMs. Context windows are still tiny and the codebases that actually run our world are still huge.

This time next year I expect you're going to be disappointed.

Let me rephrase that. This time next year I expect that I will be disappointed but you will still be predicting a completely different universe... next year.

> Keep telling yourself this. It's only hurting you.

I don't tell myself this. I just read your comment.

You're welcome to be hired in my team, we're hiring. Then you can sit there and ask claude or copilot and be fired when your trial period ends.

>You're welcome to be hired in my team, we're hiring. Then you can sit there and ask claude or copilot and be fired when your trial period ends.

I see this attitude a lot. As if "uses AI" === "can't program". Keep swinging that hammer, John Henry.

Reading your comment, regardless of AI I think that.

You're obviously getting a lot of pushback, do you think people here haven't used these tools?

Yes it's great at React/Frontend. Because that is mainly boilerplate. Hooking up click handlers and validators and submit buttons and fetch calls is boilerplate. It's just busy work. There can be complicated FE work, like making something like google maps, but how many of us ever need to do that? Most of the time if you did need to do something complicated, you get a library to do it ALL for you. Charts, maps, etc.

If you've worked on more niche languages/tools/problems you start seeing behind the curtain. The trick. It's essentially just copying someone else's code and replacing the variables. I've inherited some robot framework tests. I tried it to fix a problem, the code it spat out (v1) was nonsense. I told it. It spat out some different code, v2. It's still wrong. It then spat out v1 again! It seemed to only have these two "templates" to use. No matter how I framed the questions, v1, v2, v1, v2, v1, v2. They didn't even compile.

Hence why it's great at React, there are millions of examples for it to combine and refine.

I'd love AI to take away a lot of boring coding. But in my honest experience anything much longer than 100/200 lines and the AI starts screwing it up. It can't work on a full codebase. It even does incredibly irritating things like remove code that's needed, forgetting code changes it's already made, and worse still deleting code updates or fixes you've made.

... next year

Maybe... Why are you so sure? It's incredible what these models do but now it's all incremental and it's pretty clear LLMs are as bad at complexity as humans. If a system is of a realistic size of a business system, the llm will fuck up things mostly all the time unless you are a good coder with patience. It saves a lot of time, but it saves a lot of time of stuff that is 'just typing'; that is more than many programmers in this world can claim, but that's really another case. You are not doing anything that requires anything more than hello world context, because LLMs now really break down there.

I remember hearing this about self driving cars.

Don't worry; Elon says it's coming end of the year.

The large models are as good as their data, and there's a lot of data for common technologies and tasks, i.e. web stuff.

Step out of that bubble, though, and it's a different world.

I am aware, I just don't see this. I too write a lot of code like this, but the LLMs cannot figure out many things. If you are on that track, it will sometimes just loop forever. But sure, most humans are pretty bad too so many fail at the same things the llms do and then the llms are a lot cheaper.

Most of the time, yes. But there are certain times where I prompt an AI to build X, and it delivers a seriously neat project. It's not all of the time, though.

It has only happened like twice where I saw the full power of LLMs. I don't understand it. It's like there's a switch to dumb these things down, but if you catch them at the right time, they will blow your mind. But yes, most times, they are horrible.

Or maybe I'm just crazy, i dunno.

I think he writes HTML pages and some JS occasionally.

Frontend is a low bar for anyone and is the exact job which will be wiped out by AI. You are trapped in your own skillset prison that you don't understand what more complex engineering looks like.

My experience is that around 200-300 lines, the LLM is no longer able to understand the context of its own work. At that point, it also usually contains a significant mistake.

And around that scale, you’ll get much better code if you refactor it yourself and clean up its logic (which tends to be verbose, at best).

Someday, it will be better; today, none of them (Grok, ChatGPT, Claude) are good enough for what you describe.

>Someday, it will be better; today, none of them (Grok, ChatGPT, Claude) are good enough for what you describe.

You're living in 2023. Try Cline with Sonnet 3.7.

How big is your project (LOC)? How big are your individual source files? What language(s)?