I get that they are trying to use the human responses as a training set.
But the scope of these general purpose concierge services mean that what they are going after is damn near Artificial General Intelligence. At least in terms of capability, after-all ordering a burrito and locating an antique skull is a pretty broad capability range.
And at what point of accuracy are they aiming for for the systems to run on their own?
I guess they're thinking they can get pretty good at identifying the easy requests the system can fulfill at 99.9999% accuracy and then shift the harder requests to a human.
At that point it seems like they're just building a system that can distinguish hard requests from easy requests. Sure fulfilling the easy requests with AI is a feat in itself, but how many of those easy requests does it make sense to go through a third party? Why doesn't the burrito shop just set up it's own ordering bot?
Am I missing something?
The edge cases seem so broad that to sufficiently fulfill them would require something like AGI and the more approachable tasks of ordering pizza or burritos seem unnecessary of a third party.
I've said it before[1] and I'll say it again...the surge of interest in chatbots is premature. We're not even close to solving this problem, because yeah, it has to "learn" and "understand" language, which is equivalent to full intelligence according to some theories.
My guess is, for the foreseeable future, these services will fall into one of two categories:
1. Backed by legions of humans, and thus not that
interesting.
2. Trivial at best (ordering pizza with marginally less effort than before, a glorified collection of Slack plugins), disastrous at worst (Microsoft's Tay, the subject of my other post).
I guess one answer to my question might be:
They are trying to build an AGI.
But one would have to say that the relatively sparse training data consisting mostly of takeout orders, and the sweatshop like environment would be a strange way to go about attempting such a thing.
But the scope of these general purpose concierge services mean that what they are going after is damn near Artificial General Intelligence. At least in terms of capability, after-all ordering a burrito and locating an antique skull is a pretty broad capability range.
And at what point of accuracy are they aiming for for the systems to run on their own?
I guess they're thinking they can get pretty good at identifying the easy requests the system can fulfill at 99.9999% accuracy and then shift the harder requests to a human.
At that point it seems like they're just building a system that can distinguish hard requests from easy requests. Sure fulfilling the easy requests with AI is a feat in itself, but how many of those easy requests does it make sense to go through a third party? Why doesn't the burrito shop just set up it's own ordering bot?
Am I missing something?
The edge cases seem so broad that to sufficiently fulfill them would require something like AGI and the more approachable tasks of ordering pizza or burritos seem unnecessary of a third party.