Cheaper? BS, the RAM of your PC cost cents to make. If you have the RAM online, you'll pay for every minute using it + your ISP for the bandwith, and they can leave you with an paperweight instead of an PC at any moment they want. Wrong, not even an PC, it's only a data terminal you own and pay for, less as an smart TV. Centralizing Internet even more in the hands of few billonair pedo-assholes and their AIs, with Datacenters in former Woods (see those in Finnland by Google), what can go wrong?

The Altman User Support Center

You had me up until “cheaper because you only pay for the time that you are using it”.

We don’t really have a good benchmark there, and we’ve seen that tokens are actually much more expensive than companies are leading on as they’re trying to get people hooked. As with on-premise, or anything you own, your usage can far exceed a rented rate and then becomes cheaper and cheaper each use.

I’d be curious if there are any rented mediums that are actually cheaper once you start using them a decent amount? Cars, cloud, video, all of those become more expensive when rented if used too much.

I think it might not be. Of course, what Sam means is "let us have all the RAM so you can run a ChatGPT client on your Commodore 64", and if that were the only thing RAM was used for, that would be more efficient. But it's not and it would be undesirable for things to become like that.

So let's imagine a general purpose RAM collective. I think for that to be workable at all, it would require a major upgrade to internet infrastructure to make it sufficiently high-throughput and low-latency, which would be very costly and also require a lot of RAM. The cost of all that would probably dwarf the inefficiencies of individually owned RAM not being used maximally.

Additionally, to get the latency low enough, the data centers would have to be pretty physically close to the people using them. But people who live near each other also tend to sleep, work and play at similar times, so the gained efficiency wouldn't be nearly as high as it could be. And on top of all that, even the lowest achievable latency would still be very slow by RAM standards, so at that point you might as well be using a hard drive as working memory.