The "intelligence API" layer seems like a pretty rough business to be in. The smarter the models the more they're commodities (since they can use any harness or scaffolding) resulting in a race to the bottom on pricing. And then each new better model requires double the cost of the previous one.
I'm surprised there's not a bigger push into better and useful memory, since that would result in actual lock-in.
I'm surprised there's not a bigger push into better and useful memory, since that would result in actual lock-in.