This is an excellent point. I think the memo is making a different kind of case though - it's saying that large multipurpose models don't matter because people already have the ability to get better performance on the problems they actually care about from isolated training. It's kind of a PC-vs-datacenter argument, or, to bring it back to Maps, it'd be like saying mapping the world is pointless because what interests people is only their neighborhood.

I don't buy this for Maps, but it's worth highlighting that this isn't the usual "community supported stuff will eat commercial stuff once it gets to critical mass" type of argument.