Comment by visarga
5 days ago
> Frontier AI companies are selling at a loss.
There are huge economies to be had by batching requests and using lots of RAM for MoE (sparse models). You can't achieve that efficiency at batch size 1 on a single node.
5 days ago
> Frontier AI companies are selling at a loss.
There are huge economies to be had by batching requests and using lots of RAM for MoE (sparse models). You can't achieve that efficiency at batch size 1 on a single node.
Exactly, they put a lot of money into engineering and it does give results
Inference isn't the problem.