“somehow they’ve made 10tf = 12tf with alot less cores”
We should wait until the generation is out of it’s infancy before making such claims. Everyone knows 12TF compute will outstrip 10TF, the question is will devs make use of it – especially 3rd party devs.
As for can the PS5 do ML – you can in theory run ML on any device, just a machine with dedicated cores will be able to do it faster/on a larger scale. This usage case isn’t really big enough to tell how well it would scale across more intensive use of ML.