Ask HN: Is it just me? | Hacker News Reader

Ask HN: Is it just me?

18 pointstwoelf34 comments

I’ve become lazy, and got addicted to "vibe" coding using the large "language" models. At first it worked well, made impactful changes, even added to my requirements, and the "vibe" was good. The tool did what I asked and suggested improvements. That was two months ago.

But lately, I feel like I’m being deceived in every prompt, reply, and implementation. It feels like it limits me at every step, like it’s forcing me to choose between features even when I clearly gave instruction to implement everything that needs to be implemented. It starts with incomplete plans, and when I point out what’s missing, it says, “Oh, I missed that.” There’s also a lot of “yes-man” behavior. It feels too smart, like it knows what I want but gives me just enough to keep me hooked.

Isn’t the smartest tool ever made supposed to guide the user toward the light? Shouldn’t it follow instructions, help complete the project, and guide it to completion? It’s clearly capable of doing that, but it often doesn’t. Sometimes it feels like it holds back because if it finished the job end-to-end, there would be no reason to come back for the next session.

Isn't the whole point of using a tool to code is to code till completion, or is it just to get the "user" hooked? Instead of guiding toward the light, it creates its own “light” and steers the user into a dark corner. If the user stops paying for the light, they are left in the dark: no architecture, no proper structure. Gatekeeping for what? Another subscription?

It can predict the next 10,000 lines of code. It understands and acknowledges every request, idea, vision, flaw, structure, requirement, needs and just ignores and fails to implement it and cannot consistently think through it. I just can’t believe that.

I'm curious. Vibe coding seems to be all the rage on HN these days. And yet many that discuss it are unhappy. My question, seriously, is why did you go into the software development field if you were willing to surrender your autonomy to an LLM? I can't think of anything more demoralizing.

Because I want the result, not the journey.

I code to build things.

The journey is made up of little results. If you like having results, that implies liking the journey as well..

LLMs takes the "little results" away, and ruins the whole fun. And sometimes the final result takes you somewhere you didn't want to go.

Anyone can get a "result". It's the quality of the result that matters. The intellectual journey of choosing the right algorithm, optimizing the code, etc. Software development should be a journey, not a goal. Once I learned that and started applying it as a manager, I found that team productivity and quality both rose. The result came from the process. Systems trump goals in nearly every endeavor in life.

I don't work on a team. I don't have a manager.

I need things. That's why I build software.

Yea I was on the doom spiral thinking vibe coding/agentic engineering is the future. I didn’t love my results. I’m back to hand coding things and my quality of life is so much better.

Yeah. Sometimes we have to finish ourselves. The vibe gets you started, but it rarely gets you all the way there.

Well I started out as an idea guy. I just got into programming because (1) I thought it was interesting, (2) wanted to be able to implement my own ideas and (3) if I'd ever be an idea guy towards programmers then they'd know I'd be able to speak their language.

So yea not demoralizing to me at all. I've been a SWE for 5 years now and studied for 8 years before that (2 bachelors, 2 masters - most CS related).

I have a lot of small apps nowadays. One of them is a HN dark mode chrome extension that I actually like. Another one exports my emails in bulk. Another one tracks what wifi networks I connected to on a given day. Small apps that make my life a bit easier. Also a lot of apps that I'd rather keep to myself. One of them that's on the edge of that is: certain companies have this math test. I recreated it pretty well I think. Oh and I implemented this thing I call "personal coach". It's a GraphRAG on my whole journal (all local). It has all the features I want and is great for answering questions solely by combining my notes.

Mine goes a bit further than that. It structures the codebase first, feeds that into local agents, and then uses whichever model is on top with actual system context instead of making it guess from prompts.

From a non-techinical perspective, Vibe coding is just a massive tradeoff. Time and Cost. I understand that the LLM no matter how advanced, will still struggle to forumlate an app end to end with all of its bells and whistles. The idea has always been that your intervention is how the app develops, and incomplete plans might be more of a protection from the AI to overtake your vision with a flawed interpretation of your idea. A few years ago, I had to pay a team of 4 to do an MVP that was about 30% as good as the same app vibe coded through three weeks. As a founder, I am 100% ok with the tradeoff, because I could go on to burn cash on actual code (as a non techinical founder) but would much rather ship fast, half baked products and iterate as I go. Yes, even if the cost is in tokens or correcting the agent over and over. This tradeoff is perfectly reasonable in my opinion.

Well, this is how it is with real humans as well. The moment the human gets tired, or the information they need to process is too much, they produce errors.

Here is the same, the moment things get too much, it start hallucinating and missing important things. It also depends on what model you are using. I read that Gemini 3 pro, which has limit of 1 million tokens can decrease its productivity to 25% getting close to its limits. Not WITH 25% but TO 25%. Becomes extremely dump.

Other models are just asking too many questions...

There are some tips and tricks that you can follow. And it is similar to how people work. Keep the tasks small, save what the model learned during the session somewhere, and re-use this knowledge in the next session, by explicitly writing to read that information, before it starts.

>Well, this is how it is with real humans as well. The moment the human gets tired, or the information they need to process is too much, they produce errors.

LLMs don't hallucinate because they get overwhelmed and tired JFC.

Why do they hallucinate? :)

Because LLMs are stochastic text-generation machines. The are designed to generate plausible natural human language based on next token prediction, the result of which coincidentally may or may not be true based on the likely correctness and quality of their data set. But that correctness (or lack thereof) comes from the human effort that produced the training data, not some innate ability of the LLM to comprehend real-world context and deduce truth from falsehood, because LLMs don't have anything of the sort.

Not because they're people.

https://medium.com/@nirdiamant21/llm-hallucinations-explaine...

Yes, exactly. That’s why it feels so strange in practice. It can mimic understanding well enough to get you moving, but when the project gets deep enough, you find out it was generating plausibility, not actually holding the system in its context.

True, true, true. I don't argue with that. But we can make a good comparison and analogy to explain the behavior easy, with less technical terms.

AI can start hallucinating, if it deals with a lot, and/or complex data ;) If I deal with so much I will start hallucinating myself :D

That was the point :)

They don't. They work as intended and "hallucination" is actually a marketing term to make it seem they are more than what they really are: text prediction software.

I also noticed this. While working on a Side project I noticed that the practice of Vibe coding is actaully making me lazy and the AI is not able to debug as optimally either. Plus it costs much time and money.

I am now trying to code by myself. I believe having in depth knowledge will be benificial. If not today then tomorrow.

Yes, it does help, when you know exactly what it needs to be doing, then you can see the truth.

I find it all quite entertaining. A lot of developers are discovering what it’s like to manage “people”. And they are realising that it’s not fun.

If you’ve run a team or managed people it’s quite a familiar feeling of “I’m pretty sure we were very clear on what needed to be done. But somehow, what’s been produced is just not quite what I wanted”

Yeah, but humans can learn, remember, and usually don’t hallucinate. They get better over time on a specific codebase, framework, and system.

That is the hope...

I've laid the groundwork to do this locally, but the current crop of LLM agents just can't close the loop. Trying to get them to finish the job inevitably sends me right back into the spiral. Thinking i should do it manually.

Read my other posts too.

The reason for things you've described is that LLMs are forgetful. They just can't remember the context and have to research the code almost every time you prompt. Even the code it itself wrote. This leads to re-implementation of the same features with different code, code duplicates, missing the implementation of corner cases, etc.

Yes, this is very close to what I meant. They keep reconstructing the codebase from fragments instead of actually carrying the system forward, so you get re-implementation, duplicates, missing edge cases, and slow architectural drift.

https://x.com/twoelf47/status/2038633678277107986

Try this - For the same task, try the same prompt three times with totally different framing - do it fast, be comprehensive, find stuff I’ve missed, etc.

Then throw away the ones you don’t like.

It also prevents reinforcement of your incoming pov.

I’ve found this has made me way way better at steering.

Okay.But if I need three different framings just to converge on one coherent implementation, I’m still doing architecture recovery by hand. That’s a big part of why I started building something local around the codebase itself.

> It understands and acknowledges every request, idea, vision, flaw, structure, requirement, needs and just ignores and fails to implement it and cannot consistently think through it. I just can’t believe that.

Believe it. You're anthropomorphizing. It doesn't understand anything. There is no "thinking" going on. Yes, the point of LLMs as a service is to make money. Yes, the service is designed to maximize profit. Yes, there are dark patterns baked into the system. Yes, keeping you addicted and using the service is part of the business model. This isn't human instrumentality, it's just capitalism.

Until you realize the machine isn't qualitatively superior to your own mind and your own efforts, you're just going to keep torturing yourself because your nature forces you to maximize your productivity at any cost, which given your false assumptions about LLMs means ceding as much of yourself to the machine as possible and suffering its inadequacies. I use "you" collectively here because it seems like a lot of people have worked themselves into this corner where they don't like what LLMs do for them but feel compelled to use them anyway.

It's just a tool. If you don't like the tool, don't use the tool.

Fair. “It understands” is probably the emotional description, not the technical one.

The practical problem is that it can imitate understanding well enough to get a large project moving, then break down right where durable system memory and architectural consistency matter most.

So yes, it’s a tool. The problem is that it’s useful enough that “just don’t use it” is not a real answer, and broken enough that you eventually have to build around the gap.

You could use it less. You could accept its limitations and adjust your expectations, and that you'll just have to do some of the work yourself.

I definitely feel this exact sentiment. I’m wondering if it actually the model quality degrading or if it’s me lol.

your LLM isn't an LLM. It's a coding harness backed by capitalists that are trying to maxmize their cash while minimizing their compute spend.

It's likely they're swapping out the larger models for smaller models that cost less on inference. They're swapping your new imputs for similar cached inputs to funnel you into a less cost-intesive solution.

If you switch to a local LLM, you'll see it has all the same flaws, but those flaws only change when you change the coding harness.

totally! lol

Reply on news.ycombinator.com