[voice input] -> [action]
but to do more complex interactions it has to be more like [input 1] -> [ask for clarification] -> [input 2] -> [propose action, ask for confirmation ] -> [action]
from the viewpoint of developing a training set this is exponentially more complicated. Maybe it is like playing chess or learning Atari games by reinforcement learning, but those games involve a small number of elements that combine in complex ways, whereas a conversation involves a space of millions of concepts. (e.g. "20 questions")It's not so much that they need any specific feature ("buy toothpaste") that they need some way to recover the cost, for instance being more tied to the Amazon ecosystem, more likely to subscribe to Prime, etc.
It's not easy though. With roughly 70 million users they would need to generate about $150 a user in profit to cover $10 billion. At 2% margin they'd need to sell $7500 worth of stuff that they wouldn't otherwise sell.
What I don't get is where the $10 billion went: how much of it subsidizing the device? (If they sold 10 million devices last year it would be a $1000 subsidy which is incredible.) are they paying AWS prices to run the backend service? how much are they spending in software development, model training, etc?