That's a fair question. Am I sure? No. The problem is that nobody is publishing explicitly about how expensive this all truly is (which is concerning on its own). But a lot of points are out in that direction. I think that "cost of inference" may be a fake metric on its own that it's designed to obscure that you need multiple sub-queries to run one "thinking" query - but granted, I'm out of my depth, so I may be wrong here.

AI companies are running two arms races here: one: who can create a better model. two: who can create shitteir cheeper psudo-model that can pass as "AI" (which are also too expensive).

When you complain that those models can do anything useful, and it makes terrible mistakes, they reply that you don't use the truly smart one. When you complain that there is no way that the smart one will ever be scalable because it's crazy expensive, they tell you, "but look how much cheaper the 'fast' ones are". Never truly addressing the questions, and bundling those so you are never sure which one youare using.

Does anyone successfully run this openy avaible models locally on their own machines? I'm seriously asking, since, as I unserstand that doesn't work out as well as one would hope. That suggests that the resources required are still massive.