@chiefaioffice
With better chips, training cost will only reduce.
And once model’s performance converges, the cost of training all prior models will be irrelevant.
If open-source wins, then we will end up with only a handful of models (each optimising for a given metric), or a single ensemble