All we know about neural networks, foundation models, and transformer based systems show that the abstraction layer is different than the database layer. You have the embedding space and then you have the attention hits you have query-agnostic vectors and these are completely different things so take this this explains Niklas Luhmann ’s only communication can communicate
The web is a few dozen companies, each holding a filing cabinet, each with a receptionist posted out front. He’ll read you one file at a time, but only files you can name, as fast as he cares to read, and as long as his boss allows.
Desktop apps handle this by sharing the filesystem. Internet apps don’t use files; they use databases. We need to share the database.
Even a permanent, free, generously rate-limited API wouldn’t be enough. Applications need much more meaningful access than APIs can provide.
You can only ask questions someone already thought to answer. […]
Even the right questions come back in the wrong shape. […]
You can’t join across “cabinets”. […]
You can’t index data you don’t hold. […]
