Yeah, it's the "when" in "when no one needs to work one of these jobs" that I take issue with, I think that a lot of things that we might think of as being easily automated with some domain-specific AI, are actually fraught with corner cases that we probably dismiss as trivial because we're all walking around inside the equivalent of a super super powerful general purpose AI machine, and we didn't have to invent it. I selected my examples to be things that sound entirely menial and undemanding, but that contain a fair degree of subtlety and individual interpretation.
Driving cars is probably chief among these examples. I contend that we are not approaching a world where self-driving cars are going to be a super common thing in general purpose use. They make for a really impressive tech demo, and it's very cool that we've been able to solve as much of the problem space as we have, but any driver should be able to rattle off a list of dozens of odd little situations were you have to make a very fast decision that you're going to do something a little bit strange, to avoid doing something really really dumb (like, crashing, or causing a huge jam through inaction). A non-general purpose AI will always have to err on the side of caution - throwing up its hands and falling back to just avoiding any impacts and otherwise not moving - when it doesn't understand what is happening around it. My belief is that it will find itself in that situation a lot more than you might expect.
A lot of driving is just obeying some rules that are nearly almost true, which is where a dumb AI can shine, but the devil is in the details where the rules don't apply, or where the best courses of action are to temporarily disobey the rules, or where (I would argue most commonly), the rules have to be interpreted in combination with communication with other humans (which typically involves some kind of shared, but uncodified gestural or behavioural convention).
The only way these wide problem spaces can be handled by the AIs we are capable of creating (now and certainly in the near future) is to dramatically simplify them. If we rebuilt all the roads to be much more strictly defined, isolated and controlled, and replaced 100% of the human drivers with the a uniform driving AI, we could have self-driving cars next week.
I think the same principle applies to many of the manual jobs we have in society. Certainly not all, but enough that the idea that we're going to literally decimate the labour market in the next decade, seems silly to me.
Driving cars is probably chief among these examples. I contend that we are not approaching a world where self-driving cars are going to be a super common thing in general purpose use. They make for a really impressive tech demo, and it's very cool that we've been able to solve as much of the problem space as we have, but any driver should be able to rattle off a list of dozens of odd little situations were you have to make a very fast decision that you're going to do something a little bit strange, to avoid doing something really really dumb (like, crashing, or causing a huge jam through inaction). A non-general purpose AI will always have to err on the side of caution - throwing up its hands and falling back to just avoiding any impacts and otherwise not moving - when it doesn't understand what is happening around it. My belief is that it will find itself in that situation a lot more than you might expect.
A lot of driving is just obeying some rules that are nearly almost true, which is where a dumb AI can shine, but the devil is in the details where the rules don't apply, or where the best courses of action are to temporarily disobey the rules, or where (I would argue most commonly), the rules have to be interpreted in combination with communication with other humans (which typically involves some kind of shared, but uncodified gestural or behavioural convention).
The only way these wide problem spaces can be handled by the AIs we are capable of creating (now and certainly in the near future) is to dramatically simplify them. If we rebuilt all the roads to be much more strictly defined, isolated and controlled, and replaced 100% of the human drivers with the a uniform driving AI, we could have self-driving cars next week.
I think the same principle applies to many of the manual jobs we have in society. Certainly not all, but enough that the idea that we're going to literally decimate the labour market in the next decade, seems silly to me.