Using AI to code and maintain systems.

[willhub]

I'm a bit conflicted and I don't think I've fully embraced it, I always feel like it's wrong, but at the same time I've never really been good with coding, it has never "clicked".

Not sure if it's my mind, but I struggle to memorise in debth things like code, I've never even managed to remember a single song actually, I really struggle with these sort of things, I was learning php in uni and it never clicked.

Now I was dabbling with VB in excel and managed to build an app at work to help me with calls, this was just using google search to troubleshoot issues.

Now with AI, I've been able to really 'build' some great tools for me, or rather 'direct' the construction of them.

I've built a really good budget planner that interfaces with google scripts so it stores it in a database, it's got a built in calendar, savings goals, interest calculation, looks sleek, it's something I'd have never been able to do previously.

At work the past week I've built a web page that's essentially call structure, sleek and modern looking, interactive, has built in prompts that pop up, note generatore etc.. it's about 3000 lines of code long, again, without AI I'd not be able to do it.

At home, I've managed to set up a server with Docker with various things such as Plex on there, again, something I'd have struggled to do before.

So AI has allowed me to do a lot of things I'd have not really been able to grasp prior.

Thoughts? Anyone else in a similar position and found it of great benefit.

Should I just ignore the doubt, let it rip and embrace it?

 

[mid_gen]

If you're just making things for personal use and understand that ultimately all these LLMs are doing is trying to give you a response that you find satisfactory, rather than actually have any engineering nous. Be aware that any software you make with it is likely riddled with performance issues and security holes....then it's all good. It's wonderful for throwing together rapid prototypes and even (in the hands of an experienced software engineer) the AI tools can be very useful at taking a vibe-coded AI slop application and turning it into something approaching production quality......but don't kid yourself that you're learning to code

 

[Vestas]

Its a very useful tool for activities you do once every few months/years - stuff that you'd rather not have to get back up to speed on for trivial changes (like json files for gateways/switches when the GUI isn't helpful).

I've been playing around with a Strix Halo box (128GB) at home and there's a fairly big code "quality" difference between models. The initial output may be similar but subsequent iterations/tweaks, not so much. Some produce code which you have to marvel at just because it runs and appears to "work" - it certainly makes bugger all sense when you read it Some produce surprisingly good code on iterations.

Had a good laugh using Retro Arcade - showed me (once again) why I should never ever EVER comment on game design

The genie is out of the bottle anyway for vibe coding. Give it a bit of time and maintenance issues will loom large. Then again I suspect the AI bubble will have deflated by that point.

It really is worth people's time to run some small local models (on a 10-16GB+ graphics card) to see what the models can/can't do. Not just for coding. Don't need to use Linux (athough its better for AMD) either.

 

[ne0]

I've been using sonnet claude 4 with copilot at work to help fix bugs and build issues and for some stuff it's genuinely been amazing as it's saved me time, frustration and scrolling through stack overflow lol. Other times it seems to go off on a tangent and I have to reign it back in a bit, but yeah it definitely has its uses.

 

[willhub]

I've been using sonnet claude 4 with copilot at work to help fix bugs and build issues and for some stuff it's genuinely been amazing as it's saved me time, frustration and scrolling through stack overflow lol. Other times it seems to go off on a tangent and I have to reign it back in a bit, but yeah it definitely has its uses.

Indeed, I have been using co-pilot at work, it seems more temperamental than co-pilot on my personal machines and after some time it's clear I just need to start a new chat as it's completely lost it and rather than continuing troubleshooting could go off on a tangent about making sandwiches .

 

[willhub]

@mid_gen - Indeed, I'm not under the illusion I'm learning to code, hence my line "Now with AI, I've been able to really 'build' some great tools for me, or rather 'direct' the construction of them.", but, I do think you CAN learn to code with it, if you use it specifically for this purpose, rather than build something with it instructing it, instead exploring code, triangulating this with traditional means as well.

As for performance issues, perhaps, but for a website I don't think that is much of a concern with the power of current PCs, if I was developing some sort of system or a full on application, or game programming etc... I could fully appreciate that performance considerations are top of the list. (Well given the state of a lot of games, perhaps it isn't top of the list).


@Vestas

I'm wondering if it's worth using my 4080 for a local LLM but not sure the benefits because it'll be much slower I assume?

 

[mid_gen]

@mid_gen - Indeed, I'm not under the illusion I'm learning to code, hence my line "Now with AI, I've been able to really 'build' some great tools for me, or rather 'direct' the construction of them.", but, I do think you CAN learn to code with it, if you use it specifically for this purpose, rather than build something with it instructing it, instead exploring code, triangulating this with traditional means as well.

As for performance issues, perhaps, but for a website I don't think that is much of a concern with the power of current PCs, if I was developing some sort of system or a full on application, or game programming etc... I could fully appreciate that performance considerations are top of the list. (Well given the state of a lot of games, perhaps it isn't top of the list).

For self hosting stuff at home that's never going to be publicly accessible, you've got nothing to lose really. Install docker, and Codex/Claude or whatever, and go nuts. I've recently switched to using self hosted Immich server instead of google photos/iCloud Photos. Didn't need AI for that one mind, it's a very easy docker compose file...and Tailscale for remote access.

@Vestas

I'm wondering if it's worth using my 4080 for a local LLM but not sure the benefits because it'll be much slower I assume?

The frontier models (Claude Opus/Sonnet 4.5, ChatGPT 5.1, etc) are far more powerful than anything you can run at home, and probably faster.....it can be fun to tinker though....and you don't need to pay a subscription. You can just install Ollama and run models from there, it's dead easy.

 

[Vestas]

I'm wondering if it's worth using my 4080 for a local LLM but not sure the benefits because it'll be much slower I assume?

It really depends on the model you use because memory is everything with LLMs. If it doesn't fit then it runs like ****

There's a lot of models which do fit into 16GB (or even 12-13GB) so they will run well on most of the NVidia range.

There's also plenty of software which runs on Windows so you don't even have to mess about with Docker. You will probably want to install git - again you can do that in Windows. You probably won't get the same performance as you would in Linux but its not bad these days.

On the Corsair box (Strix Halo 128GB, which is about 4070 standard for peak compute) using QWEN3-30B-A3B-GGUF I get 70.86 tokens/sec with a time to first token of 0.55s when I prompt with "using html, js and css generate a browser-based os".

That model uses 17.7GB of graphics memory.

What gets produced is a bit naff but you can iterate it quickly.

Another one which gets about the same performance is "using a raspberry pi 5 design a door access control system".

Provided you use the right models and are using it for specific tasks then context size isn't as much of a memory problem as a lot of the small models are limited to 4096.

You should get decent performance out of a 4080 with 16GB onboard provided the model fits.

Edit - those figures were under Windows (Lemonade server) because that's where it was booted when I tested. I know ROCm (AMD cards) is better on Linux and since the DGX Spark runs on Linux I guess NVidia is as well, although I dunno about consumer cards.

 

[radderfire]

Should I just ignore the doubt, let it rip and embrace it?

Simply put, yes.

 

[mid_gen]

If you're interested in trying out a coding agent on the cheap (i.e. free), the latest Mistral model is free for a limited time, with their 'Vibe' CLI coding agent.

I've been using it a bit in parallel with Codex and Claude and it's pretty decent, fast.

 

[willhub]

If you're interested in trying out a coding agent on the cheap (i.e. free), the latest Mistral model is free for a limited time, with their 'Vibe' CLI coding agent.

I've been using it a bit in parallel with Codex and Claude and it's pretty decent, fast.

Is that a local model? For local people I assume.

 

[mid_gen]

Is that a local model? For local people I assume.

You can use the full Devstral 2 model through the API for free for a limited time ( I expect it will be very limited).

They have also released an open weights model Devstral Small 2 which is scoring only a little behind the frontier models, and can run on a gpu.

I just downloaded devstral-small-2 and am running it on my 5080 at the moment, it only just squeezes on (15GB).....but runs at a very usable speed. Not sure what size context window it's got running like this mind.

 

[Tykey]

I'm a bit conflicted and I don't think I've fully embraced it, I always feel like it's wrong, but at the same time I've never really been good with coding, it has never "clicked".

Not sure if it's my mind, but I struggle to memorise in debth things like code, I've never even managed to remember a single song actually, I really struggle with these sort of things, I was learning php in uni and it never clicked.

Now I was dabbling with VB in excel and managed to build an app at work to help me with calls, this was just using google search to troubleshoot issues.

Now with AI, I've been able to really 'build' some great tools for me, or rather 'direct' the construction of them.

I've built a really good budget planner that interfaces with google scripts so it stores it in a database, it's got a built in calendar, savings goals, interest calculation, looks sleek, it's something I'd have never been able to do previously.

At work the past week I've built a web page that's essentially call structure, sleek and modern looking, interactive, has built in prompts that pop up, note generatore etc.. it's about 3000 lines of code long, again, without AI I'd not be able to do it.

At home, I've managed to set up a server with Docker with various things such as Plex on there, again, something I'd have struggled to do before.

So AI has allowed me to do a lot of things I'd have not really been able to grasp prior.

Thoughts? Anyone else in a similar position and found it of great benefit.

Should I just ignore the doubt, let it rip and embrace it?

Good for you OP!

You can learn to code as you're vibecoding, if you choose to. In your prompts include asking the AI to explain the basics of the code it outputs. Also, I strongly advise you to do a bit of learning on the basics of coding in the languages you're using, just things like what variables, methods, classes, etc are. No need to go into things like polymorphism or encapsulation. Just spend some time watching some YouTube vids and having a little go at what's being shown.

I use AI for help in coding and you'll find most devs do, even the 1's that dismiss it, while later revealing they let multiple agents run autonomously in their project! I'm not a vibecoder but I've no problem with it, I've coded off and on for years and don't let the AI do anything more than highly specific modularised pieces of coding for me. I am aware of everything in my project. As long as you're not integrating payment systems or dealing with sensitive data etc or expecting a big number of users, there's no reason why you can't build a publically available website.

If it fails or under load etc it doesn't really matter, no harm done, just rebuild it better, you'll learn from it. There's quite a lot of animosity from some experienced devs towards vibecoders, some of it deserved when the vibecoder makes exaggerated claims etc, but a lot of it undeserved and imo pompous, it's gatekeeping imo, you seem realistic in your approach, hang on to that and try to learn about the code as you go, just the basic structure and basic concepts.

AI has opened the door to development for so many people which is great imo. It will probably make a lot developers without work eventually but I think we're still some way off from that yet and even then devs will just have to evolve their skillset.

Keep up the good work!

 

[Vestas]

Not sure what size context window it's got running like this mind.

256k apparently.