I think that hardware will adapt to serve the need. Might not look like AI accelerator cards like 3D accelerator cards in the 90s but I think compute will be ca

I think that hardware will adapt to serve the need. Might not look like AI accelerator cards like 3D accelerator cards in the 90s but I think compute will be carved out just to do on device llms. 7B param models are already really fast on CPU and memory on a pretty normal system, you could probably come up with a technique to do good character dialog with a 7B model.