@alex@eloquer.de Yes it matters for many reasons. First, neuronal networks degenerate without new input. And once they are trained on 99% LLM output, it just takes 10 generations to get garbage out on the level of a child's first words. That is especially true if future models are bound to one or two companies, because nobody else can pay the costs of development or the hardware to run on. That also causes heavily democratic issues. Furthermore the models will be manipulated more and more over time -> enshittification. How can you be sure that it creates the program you want without build-in tracking? And do we have 100% green energy in 50 years (which we require much earlier) or is LLM one of the reasons we could not reach it? In best case it is a bet that "everything will just go right", but the experience in tech with super scaled companies is more that everything will go wrong that could go wrong. What if one of the companies becomes so big, that one or two produce 95% of all software in the world via LLM? It is a highly political topic, not just for "is LLM output still FOSS compatible or not?" (which is not as far as the LLM itself is not 100% FOSS, which is far more than "open weight").