I always imagine that these services are supposed to serve as the "automation put me out of a job" of the full-time personal assistants had by the wealthy.
In that role, it's usually the assistant's job to present a selection that's sorted by optimality on axes of both quality and price (and arrival time, reliability, packaging options, etc.) Properly done, this means that you should almost always be "defaulting" to the first option presented; rarely switching to the second or third (and there only really needs to be two or three) when there's something in particular that stands out about them despite them optimizing lower.
When described this way, it's clear that this is a much harder job than just doing fancy language recognition. An AI personal assistant has to know what you care about—what weights you put on various optimization criteria—without you having to explicitly specify them. Presumably it would be an online learning system, and screw up a bit at first.
I think the key in such systems is how they'd behave before being trained by lots of individual preference data over time, though. Hopefully it could "guess" at some initial preferences in a good way. (Maybe by taking all the preferences of every other user, putting your sparse preference-set as a point in that preference space, and then seeing how you cluster: effectively "stereotyping" you based on its other "employers.")
On another tangent, one thing I haven't seen built into any such system yet is the "active academy model" mentioned here (https://scifiinterfaces.wordpress.com/2014/06/24/course-opti...). One of the criteria you'd expect a PA to optimize for in their employer is patience. Rather than staying with an initial optimization-problem solution, if the answer is not time-sensitive, one can continue to search the solution-space for new, better solutions that might crop up (involving things that maybe didn't even exist at the time the question was first asked.) If given two weeks to book a hotel room, there might be a good hotel room on the market now, but a better one on the market later. There are very complex questions involving the costs of reserving and then cancelling reservations to re-book, or holding off on reserving and maybe never finding anything better and running out of time.
For physical goods, there is even the possibility of an "indefinite optimization problem"—if I tell my PA to "buy me a top-of-the-line computer", then presumably I want my computer to continue being top-of-the-line unless otherwise noted, and this will require constant weighings of various components' or configurations' depreciation-rate, market-liquidity, employer's opportunity-cost for maintenance time, etc. All the questions a major corporate IT "buyer" employee considers, behind the facade of a single innocent request.
In that role, it's usually the assistant's job to present a selection that's sorted by optimality on axes of both quality and price (and arrival time, reliability, packaging options, etc.) Properly done, this means that you should almost always be "defaulting" to the first option presented; rarely switching to the second or third (and there only really needs to be two or three) when there's something in particular that stands out about them despite them optimizing lower.
When described this way, it's clear that this is a much harder job than just doing fancy language recognition. An AI personal assistant has to know what you care about—what weights you put on various optimization criteria—without you having to explicitly specify them. Presumably it would be an online learning system, and screw up a bit at first.
I think the key in such systems is how they'd behave before being trained by lots of individual preference data over time, though. Hopefully it could "guess" at some initial preferences in a good way. (Maybe by taking all the preferences of every other user, putting your sparse preference-set as a point in that preference space, and then seeing how you cluster: effectively "stereotyping" you based on its other "employers.")
On another tangent, one thing I haven't seen built into any such system yet is the "active academy model" mentioned here (https://scifiinterfaces.wordpress.com/2014/06/24/course-opti...). One of the criteria you'd expect a PA to optimize for in their employer is patience. Rather than staying with an initial optimization-problem solution, if the answer is not time-sensitive, one can continue to search the solution-space for new, better solutions that might crop up (involving things that maybe didn't even exist at the time the question was first asked.) If given two weeks to book a hotel room, there might be a good hotel room on the market now, but a better one on the market later. There are very complex questions involving the costs of reserving and then cancelling reservations to re-book, or holding off on reserving and maybe never finding anything better and running out of time.
For physical goods, there is even the possibility of an "indefinite optimization problem"—if I tell my PA to "buy me a top-of-the-line computer", then presumably I want my computer to continue being top-of-the-line unless otherwise noted, and this will require constant weighings of various components' or configurations' depreciation-rate, market-liquidity, employer's opportunity-cost for maintenance time, etc. All the questions a major corporate IT "buyer" employee considers, behind the facade of a single innocent request.