Would this work? I can't see meaningful work being down on the dataset piecewise.

Don't you need the load the entire parameter set to backpropagate and potentially at quite high precision. Then probably load less than 8gb as typical max vram. And then finally send that all back.

I think you really need dedicated data centres if only to move the parameters around and load the entire thing at once.

Or we could just...pay for stuff. My dear mother isn't going to spin up a GPU so she can talk to her ai assistant. I don't want to be told that I can't do stuff on my phone because I need to purchase a second stack of GPUs for my home because my kids talked to Elmo too much or whatever. We can make data centers that will gladly take money in exchange for their services, and I'm happy to provide money for them to supply those services.

ACT includes a proof-of-compute mechanism to ensure the reliability of the resources on the decentralized network. Users can register their instances with the system and be rewarded with ACT tokens for doing so. The system will also continuously monitor the instances, and remove instances that fail to meet quality standards or provide inaccurate results. ACT also includes a reputation system that allows holders to track the performance of the resources on the decentralized network. Users that provide accurate and stable resources will have a higher reputation score and be more likely to be selected for providing compute resources. The decentralized network is managed by a process for adding and removing resources. This ensures that the network is always providing high-quality resources and maintain the stability of the ACU value.

https://docs.google.com/document/d/1NTsnUVRzBfK6y7WNX2JURogD...