It’s why I am focused on how this kind of technology gets “offline” and into locally run models. If I can buy a top of the line MAC in a few years time with something in the vicinity of half a terabyte of unified ram, and run this kind of model on my own hardware, it would be a powerful force multiplier, an a genuinely empowering tool… as opposed to having to rent it from someone in order to remain competitive with everyone else forced to rent it.
Even then you actually lack the computing power and data to create these models in the first place. When this power is centralized it can essentially control what goes into the models output.
