Fair point. But is it even more of a ADHD amplifier then a modern smartphone connected to social media?

Sounds less then ideal. I honestly never used a current model of those big AI providers, except maybe a bit of ChatGPT 3.5 back in the day. My fiancé uses Claude is mostly satisfied. I played around with opencode and a local model and I‘m mostly impressed. I use it do admin my Linux gaming pc and to teach me Linux.

The local Qwen 3.6 works quite well, does good online research and actually proved me wrong when I thought it hallucinated some fact. Of course it’s no „real“ AI, but it’s useful.

Strange. You should think that’s a pretty basic task. Like you said, limited problem space. Every current LLM that is big enough and has the necessary guardrails and instructions should be able to handle it.

Interesting. In this case not only the free tier house LLM that’s included with with opencode is smarter, also the local model I run on 16GB graphics card can beat Gemini, lol.

Honestly, you’re example could already be done today by a local model like like Qwen 3.6 27b. There’s no need to run an expensive cloud model for such a simple task.

Wouldn’t even destroy jobs, there are always people needed to fry the burger.

Now do I want to talk Rona fucking robot while ordering burger? Hell no. But it could be done economically without any problems.

No offense, but when did you last use a LLM? Two years ago?

Granted they‘re talkative, but that’s it what they are, literal blah machines.

I mean fuck Altman and the rest of the tech bros, can’t wait until their bubble burst and they all crash. But the technology is going to stay, like it or not.

Sure! How much experience do you have with LLMs?

I have a Radeon RX 7800 XT.

Qwen 3.5-9b is blazingly fast on it. However while it’s its impressive for its size, it has its limitations. Complex tasks with several steps are too much for it.

So now I run the 3.6-35B model with llama.cpp It’s too big for my VRAM so I had to split it: everything that doesn’t fit on the graphics’s card runs in the normal RAM. That slows everything down, but with the right flags I get a bit over 20 tokens/s.

If you have problems with speed and you’re using ollama I would replace it with something faster like llama.cpp.

I recommend Qwen3.6, either the 27B dense or the 35B MoE model. Both outstanding for local models.

They can also make you smarter if you use them right. Key is to use local models and not giving the techbros any money.