Hacker News Reader

Regardless of the minimal time with LLMs, I think he hit major points on importance of clarity of abstractions, unreliability, shipping more features and working harder than even and losing touch with the underlying implementation.

If you are vibecoding an app without talking to anyone, those are problems.

This is still missing the point that LLM use isn't a binary choice between YOLO vibecoding or complete abstinence from LLM use.

Yeah there has to be a middle ground. Agentic engineering you could say is that middle ground, you know your code base and inspect it often. Isn't vibe coding treating your code as a black box where your prompts are the only interface? Sometimes I vibe code if the task is simple, if I'm doing serious work then I'm inspecting the code often. I don't think I will ever do "pure" coding ever again unless I'm in an env that is too secure to trust an agent to exist there.

I'm using Github copilot and I ran out of requests before the end of the month; this happens from time to time. But last month was the first time I decided to try the cheap models that were still accessible to me just to see what they were capable of. They're dumb as rocks.

I just don't know how many people have an overly negative opinion on AI assisted coding because they've just used the poor versions of these products given out for cheap/free. A similar critique is basing one's opinion on AI based on summary that Google provides for free in their search.

Get your company to pay for it (points to head)

Oh they do. And I could get them to pay even more but with the changes to copilot licensing, I'm not sure we will continue with it.

Is claude code any better for the value

Codex is the best value for money now in my usage.

This article comes from a niche of people who read a lot of news articles about LLMs (links scattered throughout) but have also avoided learning about the tools directly.

Like you said, the models available on free trials are usually toys compared to what developers use. Even Opus and GPT-5.5 are available on $20/month plans and you can buy a single month to try it out. The way they write about paying for a tool seeming "absurd and horrific" says it all about the level of actual research that went into their understanding. It's entirely based on news headlines.

So the free trial, designed to convince you that it's worth paying, doesn't well? That isn't the users' fault, that is the companies' fault. If a free trial sucks it's perfectly rational to not pay just in case the paid version doesn't suck.

Not free trials, free versions.

Certainly if the user gets a bad impression from the free version that's on the provider but if you're writing about it then you shouldn't get to be that ignorant.

Which models did you try? Open weight ones like Qwen and DeepSeek are getting pretty good, you just need the right harness, via OpenCode or Pi. I use Qwen 3.6 27B on my laptop with Unsloth Studio (Unsloth releases a lot of good quantizations and has great support for the latest features, recently released MTP support which can 2x token generation speed with no loss of accuracy).

I'm doing agentic coding with Claude Max, and it's like giving methamphetamine to a software developer.

When I run out of tokens, I pay for extra. It doesn't feel good, but I do it because I didn't write the codebase - the drug dealer did. Just one more "fix" and the code should be good to ship. Oh no, out of tokens again? Just one more "fix", and another.

And the code that the AI writes is sprawling and almost incomprehensibly complicated. Overly complicated. It's like a tweaker wrote it, on methamphetamine.

I can make this comparison because many years ago I once had an ex that put methamphetamine (I didn't realize they had an addiction) in one of my vitamin capsules "as a joke", and I was up for 36 hours straight writing convoluted code, and then writing voluminous notes about the code I had yet to write. I had never done that drug before, or since (why they are an ex). I don't even drink. After that episode I re-read what I had written and it was quite scattershot.

And now I get the same exact feeling when using AI to write code, or have it write tickets, or plan out something, etc.

I use these tools daily, and it's like putting a drug dealer between me and the code. Sure it writes a lot more code than I could write without it, but at what cost? I really don't like where this is headed. And I don't think most software developers using AI realize what is happening.

The future is local LLMs. So still methamphetamines but open source, free and an unlimited supply.

That won't really change anything, and in fact make the problem worse. It would be like "getting high on your own supply". Or just making meth at home so you can do it all the time. There's still a "drug dealer" in between you and the code. And you're going to have to pay $$$$ for hardware good enough to not slow you down. I've tried local models on an nVidia 4060 and it's pretty slow.

Right my example was meant to illustrate this cost benefit issue.

> I do it because I didn't write the codebase - the drug dealer did.

.... are you saying that you can't just open up a file "written" by spicy autocomplete and add/change parts of it?

If I wanted to spend a lot of time reading the code and figuring out what part actually needs changing, sure... but in this case it's faster to pay for more tokens and get the AI to do it.

For code I wrote, it's typically faster for me to modify it than to get the AI to do it.

So what you're saying is, you don't really know what spicy autocomplete is generating because you aren't reading it.

Great stuff champ. Really dispelling the idea that vibe coders have no idea what slop is being churned out. Top marks.

I still use LLMs in a "no-vibe coding" way. Essentially I use a combination of the typical auto-complete and asking it to generate tests or individual structs/classes that I then heavily modify. But no line of code goes unread and unvetted by me.

This is healthy but what about the economics - what about when the prices rise? At what point do you become more thoughtful about spend?

Open weight models are getting good.

LLM usage has costs that are open ended and rising. The author describe how he relates to that as a relentless cheapskate. This isn't supposed to be a directly applicable lesson to most, just a point of reference for further consideration. How much higher will costs go? How realistic will simple finishing off an odd idea be if the tools are charging professional rates? Much of the logic now seems to be can therefore do without much reference to costs or risks.

Talking about LLM without harness/environment engineering is like talking about children on a playground without safety measurements.

Managing agents is a lot like managing children. They will outsmart you 99% of the time. If your agentic environment isn't built for good sandboxing, you won't succeed.

If you build an environment that can represent mutual cooperation (e.g. helping an agent by doing the tool calls yourself if it was stuck) then it's soooo much better than just trying to rephrase the ruleset of the engagement.

Don't optimize for retries. Optimize for sandboxing and strong agent to agent communication that you can also observe, modify, and summarize.

People have a very strong tendency to go "current gen technology cant achieve x, therefore its impossible for this class of technology to ever achieve x"

Many many many of the problems with current gen ai will go away.

Ah, the classic "But just you wait, the next version will be better!" fallacy.

It's a fallacy because the reality is often not that…in fact sometimes the reality ends up worse. (See "enshittification", a process whereby technology gets worse over time, not better.)

The next version from that vendor. Or maybe the market commoditization.

You’re again conflating the product offerings from the capacity of the technology. It’s like kinetic energy vs potential energy.

The real wait is for the cost in disk space, memory and electricity of pretraining a transformer to be something the equivalent of a raspberry pi can accomplish. Then all the style and design choices of the few current vendors goes away.

I like that its not just a random hollow rant, it cites good personal reasons. I actually really enjoyed the post, including that Tetris study which is very interesting.

> Acting like LLM use is a spectrum between doing everything manually or handing control over to the LLM and vibecoding everything is a tired strawman argument.

But isn't the strawman here was that it wasn't a spectrum. That they couldn't just use it some, but all or nothing.

I think they meant "binary" rather than "spectrum".

Imagine being disappointed that an article is NOT clickbait :|