E.g. when I was reading up on genetic algorithms etc. 20 years ago we also expected the "revolution" to be right around the corner, and that things like genetic programming would change the world in a few years time. And while various of those methods found use some places, most places that could have used at least some of the simpler ones, still don't.
In other words, I think talking about a 3 year timeline is crazy. It's getting more attention, sure, but there is so much low-hanging fruit that most developers could be busy for the next 20 years putting in place the most trivial algoriths all over the place and we still wouldn't have picked off even the low hanging fruit where the computational resources and algorithms and data to make a big impact were well within reach 20 years ago.
This certainly means there is plenty of room for a lot developers to do very cool stuff and build careers on machine learning today, but it also mean most developers will not have to learn the state of the art - or anything near it - for a very long time.
As a concrete example I give to people, consider all of the search boxes out there on various sites - product searches, location searches, site searches - that are straight keyword based searches that don't take into account any clickstream data to improve ranking. The proportion of search boxes I see that take advantage of the available data is vanishingly small, even though very basic analysis can improve the perceived relevance of the results massively.
We certainly will see more companies invest in proper machine learning as the payoff gets higher and difficulty in taking advantage of it drops. But we will also see a huge proportion of sites that could use it continue to ignore it for years to come.
There are big business opportunities in finding ways of making a dent in that portion of the market, though, and so learning this stuff can certainly be well worth it on a personal level, but I don't believe in his timeline in terms of the overall market.