On "Vibe Coding"
tante.cc
tante.cc
My experience so far is middling; I found that it can be extremely helpful when working on novel things; greenfield projects it can lay a solid foundation for. If I'm learning some new framework it can shit out a first version that helps me understand how the thing works.
However, as soon as things get a bit more complex, they immediately start shitting the bed. Working in existing codebases of more than like 25 files and they become useless. Hell, even asking some models to document a 150 line batch install file made it hallucinate all kinds of stuff that didn't exist.
Furthermore, you really do have to talk to them like they're autistic idiot-savants. Precise steps, boundaries, everything. But even then they often just go off into the deep end. Even when I explicitly tell it to do one thing, one thing only and not change anything else, over 50% of the time it starts changing everything.
It's another tool, and like with all tools you need to know how and when to use it, I guess. As an industry we're not going to finger out the right ways if half the world refuses to use a powerdrill because well gosh darn it, hammers have always been good enough, and the other half swears up and down you don't need cement anymore because you could just as easily screw bricks together.
I gave it a really simple challenge: a note taking app. Allow the user to write a sentence of plain text, and add it to an array with a time stamp. Show the text on screen. Save it in local storage.
I could hand code this in about ten minutes.
Replit went off and built a truly impressive UI that looked like something that would take me a couple of evenings to pull together. It was professional and polished, but it didn't work. Hitting enter to add the item did nothing.
I chatted to it and explained. It confidently did a few revisions but didn't fix it.
It started adding logging and asked me to paste in the logs. After each subsequent paste it would say some variation of "ah I see the problem" and revise again.
After the third time, I didn't reply with the console logs and instead suggested we change tack. Replit completely ignored my input and carried on talking as if I was following instructions "yes I see the issue now, ok I've fixed it"
It was just a string of hallucinations, pretending to take user input but it fact not doing so. In the end I started asking it completely off topic questions but it had an imaginary convo with me instead. I gave up.
I've had much more success with non agentic LLMs like Gemini or Claude but they all are great at the initial response and slowly degrade add the conversation continues
I've noticed this too. I give the llm 3 or 4 prompts before a new session is opened and code is repasted. My work flow is a lot like
1. Here's a schema, what would you change? I assume nothing, because my schemas are perfect.
2. Given this schema, make me some structs
3. Given these structs, make me some endpoints
1.(New session) Given these endpoints that return data like this, make me a ui
1. while true: (New session), An llm wrote me this ui, what would you change?
If I try to steamroll ahead from schema to ui, I find it starts having trouble with context, like hallucinating functions it meant to write but didn't, or just straight up dropping functions it wrote before.
But that requires you to have certain levels of knowledge on that domain to begin with, which is not something you can just "vibe" your way out, at least for now.
The other two thirds can search for something else, just like the factory workers when robots took over most factory roles.
The remaining one third, are the folks doing the jobs robots can't yet do, or doing their maintenance.
Software factories are no different from physical ones, in this regard, a few iluminaries will be able to profit from AI, everyone else needs to re-train themselves.
That's what I've had most success with. Chat to create a requirements/spec document. Then convert it into a work plan. Then feed the stages into an agent for implementation. Works great and the extra "let me know which areas should be more refined" does wonders.
And they held the keys. Now that they are steadily working most of themselves out of a job, they are still stuck with their hubris-tainted glasses, unable to move on.
6 months ago I had no idea how Github works. Now I have thousands of commits and I have learnt to git reset --hard as second nature now :-) Hell, yesterday I figured out by chance, how to only revert a particular file. So yeah, baby steps.
I am "building" every day, talking to my AI Coder friend. We are having a lot of fun. I don't go full Agent Mode. I am always on Ask Mode. (Because - hard lessons - you know)
I am learning about Lighthouse (97% aggregate for my Next.js app)
Yesterday I made my site an "App", because someone asked if "there's an app for that"
Now I know about PWA and all it's requirements. Cool.
The point is. I am having fun building stuff. And that would not have been possible for me a year ago.
Ask yourself. If you as a coding genius had to develop a program without any internet access, how far would you get?
So stop crying that the bar has been lowered. Raise the bar. Be the guy that can outsmart the AI when it tries to take over the world. There will be few of you needed, but the need will be critical. We will always need a few Myron Aubs. Just not millions of them.
Happy building (Even if you are not "coding")
It just has a lot more legs to it with llms.
I'm glad you're having fun.
But the world doesn't run on prototype code. Maybe llms will help being the quality of code in the last 20 years collectively up a notch.
In other words, it would be great if llms took all that "just good enough code" we wrote in the last 2 to 3 decades and made it actually good code.
I strongly suspect this is not going to be the case. Instead, the world will be absolutely tsunamied by an avalanche of not even quite good enough code
As an engineering student transferred to applied math in the early 1980s I was able (with no real substantial internet access, literally getting floppy discs by mail, etc) to write an OS, an original commercial event tracking system for fleet management of mining vehicles, a real time marine seismic capture, storage, view and analysis system (basic PC, real time OS (that a friend wrote), Window Manager, dedicated hardware, lots of assembler, signal processing, etc).
And a whole lot more. (Bit of work on a sheep shearing robot, lot of early GIS mapping routines)
(Oh, some interesting stuff for CAYLEY/MAGMA that has recently been used to break quantuum encryption candidates .. odd just how long some fundementals last)
It seemed normal at the time to be able to do this, both a circle of friends and myself had a ball building stuff - much of it based on books, talking in person to others, and just diving in and playing about.
It's perhaps easier to get by without AI and real time always connected internet than you seem to think.
Yep, definitely would have been better if developers hadn't invented higher level languages and LLMs would be forced to output assembly.