(In fact I've been generally unimpressed by the people I've seen so far. The ML math is complicated, but there's a whole lot of "just fire this at that without understanding why" in the actual practitioners. I've found I have a better understanding of what is and is not possible and why than they seem to rather often. This comes back to my extended education background, though; my training is old but less obsolete than you may think, a lot of modern ML stuff isn't some totally new thing just invented yesterday, but a couple of slight tweaks to established stuff that worked really well combined with finally having the CPU firepower to use it.)
__________________
[1] My standard reference for this is a guy called Geoff Hinton. Some people in machine learning may have heard the name. Quoting:
AMT: Ok, so you have been working on neural networks for decades but it has only exploded in its application potential in the last few years, why is that?
Geoffrey Hinton: I think it’s mainly because of the amount of computation and the amount of data now around but it’s also partly because there have been some technical improvements in the algorithms. Particularly in the algorithms for doing unsupervised learning where you’re not told what the right answer is but the main thing is the computation and the amount of data.
http://techjaw.com/2015/06/07/geoffrey-hinton-deep-learning-...
[2] Check out the birth dates of the current generation of deep learning luminaries. Geoff Hinton: 1947. Yoshua Bengio: 1964. Yann LeCun: 1960. Jurgen Schmidhuber: 1963. It goes on. Go ahead and tell those guys they're greybeards who don't get what the kids today are doing.
I think the current ML world will have an overfitting reckoning. ML can optimize situations that are very normative, but struggles to create good solutions to situations in which the best outcomes are fairly distribute (but still specific best actions in certain scenarios).
ML cannot tell you what data or choices even matter in life, it can only help optimize those things once we've decided that. This is where age comes in as there's a correlation to, perhaps causal by, age on wisdom.
It seems like current ML (at least large-language models and transformers) will actually run out of publicly-accessible data to train on - the current models have scraped the majority of the useful text and image data on the public Internet, and it's not clear that we'll gain access to orders of magnitude more data, which according to the Chinchilla paper is the bottleneck on transformer performance: https://news.ycombinator.com/item?id=32321522
It's happening right now: the field is being flooded by data-kids who reason in different terms. They are statisticians first and programmers later. Their models can do stuff that bit-pushers like me will always struggle with. 10-15 years from now they'll run everything, and you won't be able to code a helloworld without specifying a ML model.