The problems left to solve can be tackled by an increasingly smaller percentage of the population. The guys that currently work as taxi drivers or paper pushers can't be reeducated to do cutting edge research.
Consider the low interest among scientists in spending time and effort on reproduction of previous results.
Project proposal: gather a group of former paper-pushers with Excel experience and train them up in statistics and a scientific discipline (e.g. biochemistry, medicine, astronomy). After 2-3 years of training, take say 20% of the most promising ones and give them a grant to reproduce existing studies.
This fixes both the issue of idle workforce and of poor study reproducibility. It could also turn out to be a lot of fun for the affected. Who knows, you may even end up launching a successful scientific career or two.
Of course, this would cost money, but hey automation will increase profits which can be taxed and then redirected to these kind of projects.
The problem is that a huge part of the population won't have neither the skills or the inclination to work on hard problems.
At any given time slice, skills in a given field are a pyramid. If we assume that the unsolved problems require the highest level of skill to be approached and solved, then there will _always_ be a shortage of problem-solvers (as the problems left become increasingly more difficult and complex)
So yes, it is a lack of skills.
Just to be clear, I'm not convinced there isn't, either, I just think it's an interesting enough thesis that I'm willing to consider it (sorry for no cite, I have read interesting discussions of this point, may google around for them later).
I think the basic thesis boils down to this - there are, at any given moment, an "unlimited" number of scientific and technical tasks required to solve those deep problems you mentioned. However, there are (this is a thesis, not a conclusion) a limited number of those problems that can be worked on right now. Those problems need to be solved before the next set can be solved.
Now, if you have 10 research teams working on these problems, you have a better chance of solving them in a reasonable amount of time than if only 1 research team is working on them. 100 is better than 10. 1000 is better than 100... but at some point, you're maxed out on the number of research positions that can contribute effectively. Out of 1000, 100 will take a promising approach, 10 will get it right, and... 1 will solve the problem a bit faster than the other 9. There's a point at which you hit diminishing returns.
Is this true? (and btw, I'm constructing some of this thesis myself, as you can probably tell, I don't think I've read it in quite this form before). Is it true-ish?
I'm not really sure. I'm not sure it matters for the moment, because I don't think our budget for R&D and high value engineering work, or our educational systems to develop talent, are anywhere close to a point where we should be discussing whether we're reaching diminishing returns to scale.
But is there a theoretical point at which throwing more genius at problems won't make a difference in how quickly we progress? The question interests me, in an abstract way.
That's why it's different to the old luddite fallacy. Machinery and robotic factories throughout the ages have cost millions to replace 100's of jobs. Every time you want to build a factory, it costs money.
With software that is not necessarily the case. You write it once and then distribute everywhere at little to no cost.
Automation requires really understanding the process and formalizing it. It requires knowing all possible failure points, or at least having a "A Bad Thing Has Happened" failure point of last resort and learning the details of those over time. This will annoy people no end.
The cost of bad software can be substantial.
Nothing fancy. "machine learning" has a lot of space taken up by fancy filtered "running averages" and other curve fit ... things. Then there are Kalman filters. I can't even categorically say that what I did was or was not a Kalman filter.
Once you do that, then there are degenerate and ... overflow cases that propagate like fruit flies. You keep talking about "use cases" in the presence of people for whom that is foreign.
The hard part is then accountability. One "thing" can take two smart people who know this stuff well hours to hammer out.
Never say never, but because of observer bias and path dependence, it doesn't feel like that gets any better. At some point, a GA has to admit it is powerless in the face of its own inadequacy and ask a grownup what to do. So what you're doing is distilling the set of ... opportunities for difficulty into an increasingly higher proof.
I tend to classify a lot of learning computer problems as pre- or post-general AI. As in, would it essentially take a human equivalent AI to solve this problem or can we do it sooner with shortcuts and good programming?
ie. I used to think that computer vision was completely in this post-AI field, but the recent huge strides that have been made in deep learning etc. have changed my mind considerably. Human-like ability to recognize and differentiate objects is a major step forward in replacing humans around the workplace.