Replying to
@ZephyrAuster @lritter running inferencing on small/medium models locally is quite reasonable, but *training* (and distribution) will continue to require storage and compute out of reach for many people for the foreseeable future. Since models can't "learn" during inferencing (weights are read-only), the modern commercial models are constantly getting rebuilt with updated training data.
Another problem to think about is getting access to appropriately licensed training materials either to make models reproducible or to make them "free" software where you have the ability to modify and rebuild the model that you are running.
