Because typically the resources for a particular AI job need to be centralized in order to obtain usable speeds. Even a machine with a decent 16GB GPU is going to lose speed swapping in and out of system RAM - though it does allow for larger models - so machines in a LAN can’t currently split the workloads without significant efficiency losses. It’s also why AI boards are typically using soldered LPDDR rather than swappable DIMM’s (can’t get an AI CPU with a non-soldered board).
Peers over the Internet would be abysmally slow or untenable for many workloads, at least with the way a lot of stuff is built currently.
HOWEVER, that’s assuming monolithic non-specialized/optimized models. In the future as we see models built and trained for more specific jobs, it should be possible to have individual students run more as specialists and swap data. There are models that are specialized currently but not really an ecosystem for them:
General LLM/Chat or
Image recognition
Image generation
3D Spatial navigation/manipulation
Code generation/processing
Network analysis
Audio processing/generation
Etc
If we think of something like creating an MMO (and assuming we didn’t include the moral issues of AI generation), then a pipeline could have different assets building a soundtrack, 3d models, scenes/animation per a script, character voicing, various dialog, 3d world space mechanics, etc all running from different clustered machines
Because typically the resources for a particular AI job need to be centralized in order to obtain usable speeds. Even a machine with a decent 16GB GPU is going to lose speed swapping in and out of system RAM - though it does allow for larger models - so machines in a LAN can’t currently split the workloads without significant efficiency losses. It’s also why AI boards are typically using soldered LPDDR rather than swappable DIMM’s (can’t get an AI CPU with a non-soldered board).
Peers over the Internet would be abysmally slow or untenable for many workloads, at least with the way a lot of stuff is built currently.
HOWEVER, that’s assuming monolithic non-specialized/optimized models. In the future as we see models built and trained for more specific jobs, it should be possible to have individual students run more as specialists and swap data. There are models that are specialized currently but not really an ecosystem for them:
If we think of something like creating an MMO (and assuming we didn’t include the moral issues of AI generation), then a pipeline could have different assets building a soundtrack, 3d models, scenes/animation per a script, character voicing, various dialog, 3d world space mechanics, etc all running from different clustered machines