This is easy to do badly and very hard to do well.
It's not just about inferring intent. Most discourse systems have little or no model of the consequences of their actions. Systems that can act on the user's behalf need to know if they're potentially making a big mistake or a little one, as guidance of when to ask for clarification. This requires some degree of "common sense". AI systems still suck at that.
Systems that actually do something need to be much better at this than ones that just answer questions. There needs to be some cost/benefit analysis of whether it's necessary to repeat back info and ask for confirmation. Without that, the system will either ask for confirmation too much, or screw up too much. The system needs to have a model of how clear the user is in their intent.
Pizza ordering is relatively low-risk. Travel reservations are a higher-stakes item. Medical is a long way off.
As I wrote last time, I recommend watching "The Devil Wears Prada", the scenes of Andy's first day as Miranda's personal assistant. That's the level of performance you want.
Good point. I've been thinking so much about the difficulty of inferring intent I hadn't really considered the consequences of getting it wrong. The more responsible we allow these chat bots to be the greater the risks become.
This is why I think that eventually people will bring this technology up to a useful level it's only going to be the kinds of companies that can make huge time and money investments and have access to huge datasets. Not startup territory really.
It's not just about inferring intent. Most discourse systems have little or no model of the consequences of their actions. Systems that can act on the user's behalf need to know if they're potentially making a big mistake or a little one, as guidance of when to ask for clarification. This requires some degree of "common sense". AI systems still suck at that.
Systems that actually do something need to be much better at this than ones that just answer questions. There needs to be some cost/benefit analysis of whether it's necessary to repeat back info and ask for confirmation. Without that, the system will either ask for confirmation too much, or screw up too much. The system needs to have a model of how clear the user is in their intent.
Pizza ordering is relatively low-risk. Travel reservations are a higher-stakes item. Medical is a long way off.
As I wrote last time, I recommend watching "The Devil Wears Prada", the scenes of Andy's first day as Miranda's personal assistant. That's the level of performance you want.