I've been dabbling with running AI models at home for a few years now, and I've noticed one consistent pattern among the community. Many people are impatient. On day 1 of a new model launch, it probably isn't going to work well. While the maintainers of the foundational open-source software sometimes get early access to a model, they very rarely had enough time or manpower to ensure that it is operating at 100% of its capability. In fact, it typically takes months for a model to reach its plateau, and that's assuming people remain interested. Despite this, the average forum poster will spend a couple of hours poking at a model on day 1 and declare it to be shit. What's worse is that those impatient users flood the open-source maintainers with low-quality complaints, requests, questions, etc. The irony is that just further delays the plateau.
The impatience extends beyond getting models running. It's all too common to read a comment along the lines of "this model will only run at 15 tokens per second on your hardware; I wouldn't even bother with it" or "this model thinks way too much; I just want it to get the job done." What exactly are we in such a rush for? Fifteen tokens per second isn't fast, but I can go live my life while my computer works for me. I also know that longer thinking traces are directly correlated with better performance. Why would I want the model to perform worse? If you can't be patient, then maybe local models aren't right for you.