I am not debating the transformative nature of self-driving cars or if it will become common place. I think it is huge. I guess my point is that is just one application of machine learning / AI. To automate such a complex task with our current approach to AI, that humans do easily, one will need to spend trillions of dollars and therefore the benefit, such as with self-driving cars, needs to be significant.
I suspect this will be the case with any AI workload. Heavy customization to get to a place where it is good enough. So the financial benefit needs to be equally huge and someone needs to feel confident they can extract that value 20 years in the future.
Same with computers though - back in the 60's problems had to be sufficiently valuable to solve in order to justify compute resource to them. For the initial computers, the 'prize' was Enigma and winning the war.
Now we can use compute resource for memes.
Costs will come down, so the required benefit will be lower.
I suspect this will be the case with any AI workload. Heavy customization to get to a place where it is good enough. So the financial benefit needs to be equally huge and someone needs to feel confident they can extract that value 20 years in the future.