When the released it they specifically focused on the accurate recall across the context window. There are a bunch of demos of things like giving it a whole mov

When the released it they specifically focused on the accurate recall across the context window. There are a bunch of demos of things like giving it a whole movie as input (frame every N seconds plus script or something) and asking for highly specific facts). Anecdotally, I use NotebookLM a bit, and while that’s probably RAG plus large contexts (to be clear, this is a guess not based on inside knowledge), it seems very accurate.