Perry E. Metzger
Apropos of nothing: Imagine that we had a world without the Doomers and where the expectation was fully open model weights. Could we pay for training runs? Yes, for the same reason we can pay for Linux development. Lots of organizations would probably cooperate to do big runs, including large users of models and companies whose business is selling API access to cloud inference equipment. Distributed training runs would probably involve spare cycles on machines at vast numbers of organizations. This would not happen because of communism, it would happen for the same cold hard financial reasons that large organizations do open source stuff right now. Commoditizing your complements and sharing the development expense of technologies that are not unique to your organization makes you money.