I usually use skill like openspec when requesting LLM to do something, combining them with matt's skills like grill-me. I assume both of this skill will make LLM output in certain way as I jump between model, I don't really notice them. Usually I also ask to 'talk in basic english language' or ' talk in simplified technical english language' when just asking random things in the codebase. Surely not perfect and better to have normal sounding LLM directly.
Llama 3.3 was good, tho. For the multimodal, llama 4 also use llama3.2 approach where the image and text is made into single model instead using CLIP or siglip.
They got the whole Twitter database. It's kinda the same with Gemini. But somehow Meta isn't catching up, maybe their llama 4 architecture isn't that stable to train.
Edit: Sorry, I kinda missed the point, maybe I'm sleepy when writing that comment. Yeah, I agree that LLM need big memory to run which is one of it's downside. I remember someone doing comparison that API with token based pricing is cheaper that to run it locally. But, running image generation locally is cheaper than API with step+megapixel pricing.
I usually use skill like openspec when requesting LLM to do something, combining them with matt's skills like grill-me. I assume both of this skill will make LLM output in certain way as I jump between model, I don't really notice them. Usually I also ask to 'talk in basic english language' or ' talk in simplified technical english language' when just asking random things in the codebase. Surely not perfect and better to have normal sounding LLM directly.