Next step will be to replace your gf with AI entirely.

 

Might not be a bad idea to keep it in mind they are probably asking it about you.

 

As I non-programer I've been doing simple stunts work related, like a system to ask / give feedback / organize /and publish designs for my main client that has decreased the workload on stupid admin / email exchanges to a quarter of what it was.
Also I've been tailoring all my work software with little scripts: automated renaming artboards based on client/project, automating naming exports, and basically automating everything tedious or repetitive in my personal workflow that it's linked to the above project for maximun click reduction XD.

All with local model "Devstral Small 2507 24B Q8 128k context" via https://openwebui.com/

Also I've been learning a lot of Max4Live visual programming and creating my own quirky devices.

It takes time to get results, it's not aladin's lamp. But I wouldn't imagine myself doing this stuff before the models. For simple projects, web programming/animation (making crazy gsap stuff that took me days in just minutes is magic), and non huge database stuff it's really good.

I've programer friends that uses Claude Code extensively like is a new way of doing things, they told me that they focus more on architechture and planning at the begining and then agentifies everything and debugs piece by piece. For what they told me senior programmers/software engineers + ai are being the GOAT. Newbys relaying on it without the ability to weed out the bugs are just going to not be needed. No entry level ever again. Which sucks. Unionize hard people.

Also I've been seeing around experimental projects that are creating UI interfaces for this new way of doing things, like you're playing a RTS or 4x game like Civilization: establish a goal, gather resources, builds troop (agents) and then tackles the goal. It's seems that the RTS UIs are going to be the direction to visually program in the future, which is crazy logic.

 

Torvalds has tried it for a hobby program recently and also concludes that it's suitable for fun projects at best (yet):
https://arstechnica.com/ai/2026/01/hobby-github-repo-shows-linus-torvalds-vibe-codes-sometimes/

Studies like Coding on Copilot report that code quality deteriorates with the increasing use of AI assistants due to churned code instead of good and sustainable coding solutions.

 

"Gemini - I understand your frustration! "

i get that very often.....

 

What GPU?

 

You made my weekend, thank you very much, friend!

 

I have this theory that the reason some people have been so grumpy/toxic on this forum lately is because the recent Cloudflare outages also repeatedly took down their AI girlfriends.

 

@Xupito - I found there is one thing any user MUST do if they wish to do many things in one session with Gemini, ChatGPT or CoPilot.

Assume it gets the first thing you ask for right... If you then wish to continue and have it do a similar thing, but to do a different object, or a different way of doing the same thing, you MUST tell it what not to do explicitly. In such a way it cannot possibly misinterpret what you ask for. Sometimes, starting a new session is the better way.

If you do not want it to change something you have supplied I found there are two ways to ensure it understands that:
1. "No regeneration" - this means whether images of any kind or code, you are telling it that it is not to do its own thing and add anything you did not ask for. For some reason "No post-processing" it has difficulties with regularly.
2. I found if I always start with "You do not have permission to...." and list what it is not allowed to do, the success rate is towards the highest possible outcome. "No deviation" is another thing it understands providing you have explicitly listed exactly what you want it to do.

Dealing with A.I. is like dealing with a 6-year old child with the brain of a Computing Science and Math PhD.

EDIT:

HYPOTHETICALLY:
lets say you want it to add compression a specific way to an audio file you supply and it has the capability to use every major studio effect - and you specify the ratio, attack, release, knee and threshold. In many cases, if you do not tell it that this is all it is allowed to do, there is an even chance it may decide it did not like your loudness settings, or your EQ or something else and then go ahead and add a pile of things you did not ask for.

 

Last edited: Feb 1, 2026 at 5:08 PM

You'll need at least a 3090/4090 to not fall into RAM offloads if you are going to put medium context. But there are quantized versions that you can run on a 16gb vram with some degradation (that in code isn't at all ideal, tbh, I preffer RAM offloads and wait). On a 5090/32gb runs perfect, plenty room for the context, but if you're going to use the 5090 just for this it makes no sense. Way cheaper to run it on cloud. I use it for other aspects of my work, this is a very big plus on top.

And Devstral is an european open source model, if you fancy the politics involved around software