There is nothing magical or sophisticated about AI research. Especially neural network variety. Everyone in the field knows neural network is just pure black box. All people have is a bunch of tricks to link up neural network layers in different ways and see which one works. It's not even complicated math or physics or sophisticated tech like EUV. It's just trial and error.
The only sophisticated about AI research are actually the AI chips themselves. Each AI chip and how it's manufactured is probably the most sophisticated tech in the world.
But the actual software for AI, child's play.
I don't see any reason why anyone with enough money and GPU supply cannot make progress on it.
In anything software related, iteration is fast and so the field moves fast, but the barrier of entry is low and "secret sauce" is worth much less, since it's not hard or expensive to test out different techniques. However, I disagree on it not being sophisticated. Algorithmic research is fairly close to "pure math" and no one would claim math isn't one of the most sophisticated fields in the world.
In semiconductors - and really, any hardware industry - "secret sauce" is more common because iteration speed is slow and barrier of entry is high, so companies hide their results, theoretic research moves slow, and things like optics and material sciences just need to be worked out with
time rather than brilliance. A machine has to be constructed and tested on the order of months; while an algorithm can be coded up and tested in days or even hours.
The iteration speed explains this perception that software isn't "complicated." Both fields advance by trial and error at the cutting edge, but trial and error is just that much faster and cheaper in software than hardware.
