Yes! Training costs for humans are per employee. Training costs for robots are per job function. Also robots are very stupid and very expensive to train. Automation will take over the job functions that each employ larger numbers of persons. So that will have a big economic impact, because lots of jobs are involved. But will lots of job functions be involved?
Automation could impact a few lots-of-employees job functions (such as driving cars, receptionist, and, err, my crystal ball has gone cloudy) and then stall. Researchers will have good ideas for how to automate job functions, but be unable to get the funding because only a few thousand humans do that particular job and it is cheaper to pay to train humans (times a few thousand) than it is to fund AI research (once).
There are precedents for stalls. Think about garments. The sewing machine automates the process of passing the needle through the cloth. That is a big deal and causes a dramatic step change in productivity. Then what? Not much. For a hundred years and perhaps a little longer yet garment making remains at the same level of automation, with huge numbers of persons working in factories using the same old tool.
That is why my crystal ball is cloudy. The penetration of robots into job functions that have non-huge numbers of employees depends on coming up with clever hacks analogous to the invention of the sewing machine. Without a clever trick, AI researchers might still be able to brute force things (imagine an industrial robot programmed to hold a manual needle, sewing needle-and-thread human style) but it will be too expensive and not replace human workers.
I imagine the clever tricks trickling in a few per decade, dragging out the automation of the economy over a century or two (or three).
It's not per job function though but per human function (image recognition for instance)
Keep in mind that once image recognition is done properly it's applicable to all jobs that require image recognition. Ex. a radiologist AND a quality control function.
So it's much much worse for humans ability to compete in the long and short run.
Automation could impact a few lots-of-employees job functions (such as driving cars, receptionist, and, err, my crystal ball has gone cloudy) and then stall. Researchers will have good ideas for how to automate job functions, but be unable to get the funding because only a few thousand humans do that particular job and it is cheaper to pay to train humans (times a few thousand) than it is to fund AI research (once).
There are precedents for stalls. Think about garments. The sewing machine automates the process of passing the needle through the cloth. That is a big deal and causes a dramatic step change in productivity. Then what? Not much. For a hundred years and perhaps a little longer yet garment making remains at the same level of automation, with huge numbers of persons working in factories using the same old tool.
That is why my crystal ball is cloudy. The penetration of robots into job functions that have non-huge numbers of employees depends on coming up with clever hacks analogous to the invention of the sewing machine. Without a clever trick, AI researchers might still be able to brute force things (imagine an industrial robot programmed to hold a manual needle, sewing needle-and-thread human style) but it will be too expensive and not replace human workers.
I imagine the clever tricks trickling in a few per decade, dragging out the automation of the economy over a century or two (or three).