Yep agreed

It's a mixed bag. It depends on your problem domain, the problem you are trying to solve (within that domain), the context you provide the LLM, the output it generates (you are using libraries to coerce the output into predictable (JSON) structures, right?). What's more, based on what you are trying to do, the LLM you are using might have sufficient training data, but not necessarily so (resulting in possible hallucinations/confabulations). So, there's that. Also, LLMs are not deterministic, they can (and will) generate a different response every time you call them (even if the context you provided is the same). So, yeah... sometimes these things really deliver and other times, it's just... meh!

I found the same thing, so I don't bother with ai-gen code AT ALL. I found that the time wasted fixing up the slop was not worth it - it's more efficient to code it yourself, as shown by studies (eg referred to here: https://www.linkedin.com/pulse/vibe-coding-myth-when-feeling...).

No, your vibe-coding is not more productive, unless your only metrics for productivity are commit counts, PR counts, deployment counts. I can commit, PR and deploy crap all day long and "score well" - and this is what people are clinging to with their ai-gen defenses. I'm really sorry to inform you that your experienced "speed-up" is just a trick of the brain (remembering from an article written, iirc, by Gurwinder, but I'm having trouble finding it now) - you're actually going slower, and your brain is tricking you into thinking it's faster because whilst the ai was "coding", you didn't have to, so it feels like more of a win than it actually is, considering the results.

Every single time I try to use it for research or learning it ends up spitting subtly invalid code. Results range from imaginary APIs that don't exist to straight up invalid syntax, not to mention outdated info, contradictory reasoning and flat out wrong explanations.

Maybe spending $200/mo or whatever to access the top-of-the-line models will mitigate some of that, but I'd rather come up with the solution and gain the understanding myself, then spend the money on something worthwile.

I don't use it for large-scale code generation, but I do find it useful for small code snippets. For example asking how to initialize a widget in Kendo UI with specific behavior. With snippets, I can just run the code and verify that it works with minimal effort. It's often more about reminding me of something I already knew rather than discovering something novel. I wouldn't trust it with anything novel.

In general, I think of it as a better kind of search. The knowledge available on the internet is enormous, and LLMs are pretty good at finding and synthesizing it relative to a prompt. But that's a different task than generating its own ideas. I think of it like a highly efficient secretary. I wouldn't ask my secretary how to solve a problem, but I absolutely would ask if we have any records pertaining to the problem, and perhaps would also ask for a summary of those records.

>Am I using the tools wrong or are others finding the same thing?

Like any new tool, there is a learning curve. The Curve is rather steep right now with the horizon changing to quickly. The right tool also matters a great deal; right now you can run a model at home on 32gb vram that's objectively better than gpt 3.5 from 2023 or grok 2.

>lus, the time I supposedly save building things gets eaten up debugging, correcting, improving the AI-generated slop.

Those complaining about ai slop are almost certainly complaining about lack of prompt engineering skills.

Let me also explain the proper evolution here.

In 2021, you would go to stackoverflow, copy some of your code or ask a question and hopefully someone helped you sometime. Then you'd get the help and probably paste their code in.

In 2024, you would go to AI, copy some of your code, ask a question and the AI responds quickly. The solution might be bad, buggy, and so you reprompt because your first prompt wasnt engineered well. You finally get good code and copy and paste.

In 2025, why all this copy and paste? Why not use agentic where it does the copy and paste for you. It knows what to read, and what to do.

Also 2025, what if you have AI also orchestrating 1 level higher and verifying that it itself is doing a good job.

I can't actually be bothered myself either...

Did take a look at gemini result, but then it was different from immediate results under it so didn't leave lot of confidence even to get simplest things right.

I don’t and I won’t. My large clients do not care. The day they start to require any of that nonsense, I’ll drop them as a client. Simple as that.

I have decided to be radical about AI and LLM: I don’t like them because they are a waste of time, and I would like them even less if they were this magical world-changing technology people want us to believe. I am at a point of my career where concerns of productivity or how to operate in large-scale tech companies are the least of my problems, while I increasingly appreciate the artistic craft of programming and computers, especially in small-scale to improve our lives rather than accumulate profit. So while I could admit LLMs they have their use, I want to consciously follow a path where human intelligence and well-being is of the utmost concern, and any attempt at creating intelligent machines is tantamount to blasphemy.

Under this philosophy, seeing that all the talk about imminent AGI has led to creating spam and porn at large scale, I can only roll my eyes and remain on the sidelines while we continue down this idiotic path of resource extraction and devaluation of human ingenuity for the profit of the few.

AI ruins your flow. That's the biggest problem. I sit here and wait for Claude to do something. Then I get distracted by social media.

No these things don't actually work if you study human psychology:

* Switching to another work task (what for like a minute?)

* Playing chess or something (sure its better than social media, but still a distraction)

But I do like AI tools that don't interfere with my flow like Github Copilot, or even chatting with Claude / ChatGPT about a task I'm doing.

Just started a Claude Code experiment this week. I'm building a new NAS but instead of using an off the shelf appropriate distro like TrueNAS I just installed NixOS and I'm having Claude Code fully manage the entire operating system. It's going pretty well so far. Initially it would reach for tools like dig that weren't available on the install but after a "# memorize We're on NixOS, you need to try to do things the NixOS way first. Including temporarily installing tools via nix-shell to run individual commands." those issues went away and it's doing NixOS things.

From a clean NixOS command line install, we've got containers and vms handled. Reverse proxy with cloudflare tunnels with all endpoints automatically getting and renewing SSL certs. All the *arr stack tools and other homelab stuff you'd expect. Split horizon DNS with unbound and pihole running internally. All of my configurations backed up in github. I didn't even create the cloudflare tunnels or the github repos. I had claude code handle that via API and cli tools. The last piece I'm waiting on to tie it all together are my actual data drives which should be here tomorrow.

Is this a smart thing to do? Absolutely not. Tons of things could go wrong. But NixOS is fairly resilient and rollbacks are easy. I don't actually have anything running on the NAS in use yet and I've got my synology limping along until I finish building this replacement. It's still an open question whether I'll use Claude Code like this to manage the NAS once I've migrated my data and my family has switched over. But I've had a very good experience so far.