Most of those jobs will be the hardest to automate because Moravec’s Paradox. Automation will wipe out lawyers and accountants long before it reaches sewer maintenance and refuse collection.

It seems to me, that the success of robots in manufacturing demonstrates that this does not apply to many (but not all) unskilled labor jobs.

It also seems to me that in many cases it may be possible to modify the job to fit the capabilities of a robot rather than the other way around. For example, rather than build a robot that can clean an arbitrary toilet I could either standardize the internal dimensions of the toilet bowl or give the robot a pipe-style brush that it big enough to clean any bowl.

Sure, this is an argument often made by self-driving car enthusiasts, let’s just remake all the roads, ban all pedestrians and non-self-driving vehicles and so on. But it’s another instance of the same paradox, instead of radically underestimating the computation involved in actions that nature has given us dedicated hardware for, underestimating the vast cost of doing so.

In this case not only would you need to standardise the toilet bowl, but the route through the house to get to the bathroom and what to do about any unexpected obstacles...

The paradox in general applies to everything that humans take for granted that we can do without fully appreciating that it took us a billion years of evolution to get here.

Shitty jobs are not must vulnerable, quite opposite in fact. Automating some 400K/year quant job at hedge fund makes morse sense than automatng MCjob that pays 40K/year. More to that, shitty jobs are hardest to automate (we are nowhere close to good janitor robot, for example).

Also, you can not outsource cleaning toilets or hairdessers, but you can software development or building trading models.

Also, typically lower paying jobs are physical labor jobs, trading time for money, not "thinking" jobs that deal with exceptions to the rules.

Robotics are suppose to automate lower labor physical positions, AI is suppose to replace our thinking and mental positions.

There appears to be a huge hump in difficulty of robotic “automatability” caused by the fact that humans are just so much better at these two tasks than machines are.

“Thinking” jobs that essentially entail transformation of one type of data into another seem no more difficult to automate (if not perhaps even easier) than most manual labor that has not already been automated.

My uninformed guess is that we will automate low- and mid-level “thinking” jobs and the remainder of physical jobs at around the same time, because both require near-human perceiving, communicating, and manipulating capabilities.