> AFAIK you can get the same model that they host online, it's just that there's so many distilled/quantized models floating around that the average person usin

> AFAIK you can get the same model that they host online, it's just that there's so many distilled/quantized models floating around that the average person using ollama or whatever think they're running the same thing as the one on chat.deepseek.com, but they're not. No, I was perfectly aware of this difference when writing my comment. I just don't have the hardware, time, and skill/motivation to do a scientific comparison. OpenAI used to use static weights, but it's pretty clear that they're doing more by now (and no, I don't mean just RAG). How would you go about verifying the weights used by cloud DS are the same as the ones you get when using full local DS with beefy hardware? Running millions of prompts at temp=0? > 1. how is it "monopolization", when AI companies are also providing open weights models? Please refer to the rest of my comment for my response to this. Also, AFAIK, DS was trained on ChatGPT conversations for RILHF? Regarding the licensing of training data, I feel like we are not communicating on a common basis anymore. Do you think that training needs no commercial content licensing, or do you think that most AI companies have commercial licenses (not just a paid copy!) for their training data?