

Is this correct? I was under the impresion that the most expensive part of an llm is the training, and once that’s done the cost of running a prompt is negligible.
I get your point that this last part doesn’t scale well, but the far larger cost of training must get very diluted if they distribute it across a large user base.





Yes, but don’t underestimate the power of centralisation.
6 months ago you could set up a server for running a decent local llm for under 800.
By increasing the demands and pushing the price of hardware up, they are efectibly gate keeping access to llms.
I think the plan is that we will need to rely on this companies for compute power and llms services, and then they can do all sorts of nefarious things.